Daily Papers

Federated Learning (FL) is becoming a popular paradigm for leveraging distributed data and preserving data privacy. However, due to the distributed characteristic, FL systems are vulnerable to Byzantine attacks that compromised clients attack the global model by uploading malicious model updates. With the development of layer-level and parameter-level fine-grained attacks, the attacks' stealthiness and effectiveness have been significantly improved. The existing defense mechanisms solely analyze the model-level statistics of individual model updates uploaded by clients to mitigate Byzantine attacks, which are ineffective against fine-grained attacks due to unawareness or overreaction. To address this problem, we propose SkyMask, a new attack-agnostic robust FL system that firstly leverages fine-grained learnable masks to identify malicious model updates at the parameter level. Specifically, the FL server freezes and multiplies the model updates uploaded by clients with the parameter-level masks, and trains the masks over a small clean dataset (i.e., root dataset) to learn the subtle difference between benign and malicious model updates in a high-dimension space. Our extensive experiments involve different models on three public datasets under state-of-the-art (SOTA) attacks, where the results show that SkyMask achieves up to 14% higher testing accuracy compared with SOTA defense strategies under the same attacks and successfully defends against attacks with malicious clients of a high fraction up to 80%. Code is available at https://github.com/KoalaYan/SkyMask.

Intrusion Detection Systems (IDS) play a crucial role in ensuring the security of computer networks. Machine learning has emerged as a popular approach for intrusion detection due to its ability to analyze and detect patterns in large volumes of data. However, current ML-based IDS solutions often struggle to keep pace with the ever-changing nature of attack patterns and the emergence of new attack types. Additionally, these solutions face challenges related to class imbalance, where the number of instances belonging to different classes (normal and intrusions) is significantly imbalanced, which hinders their ability to effectively detect minor classes. In this paper, we propose a novel multi-agent reinforcement learning (RL) architecture, enabling automatic, efficient, and robust network intrusion detection. To enhance the capabilities of the proposed model, we have improved the DQN algorithm by implementing the weighted mean square loss function and employing cost-sensitive learning techniques. Our solution introduces a resilient architecture designed to accommodate the addition of new attacks and effectively adapt to changes in existing attack patterns. Experimental results realized using CIC-IDS-2017 dataset, demonstrate that our approach can effectively handle the class imbalance problem and provide a fine grained classification of attacks with a very low false positive rate. In comparison to the current state-of-the-art works, our solution demonstrates a significant superiority in both detection rate and false positive rate.

Black-box adversarial attacks on Large Vision-Language Models (LVLMs) are challenging due to missing gradients and complex multimodal boundaries. While prior state-of-the-art transfer-based approaches like M-Attack perform well using local crop-level matching between source and target images, we find this induces high-variance, nearly orthogonal gradients across iterations, violating coherent local alignment and destabilizing optimization. We attribute this to (i) ViT translation sensitivity that yields spike-like gradients and (ii) structural asymmetry between source and target crops. We reformulate local matching as an asymmetric expectation over source transformations and target semantics, and build a gradient-denoising upgrade to M-Attack. On the source side, Multi-Crop Alignment (MCA) averages gradients from multiple independently sampled local views per iteration to reduce variance. On the target side, Auxiliary Target Alignment (ATA) replaces aggressive target augmentation with a small auxiliary set from a semantically correlated distribution, producing a smoother, lower-variance target manifold. We further reinterpret momentum as Patch Momentum, replaying historical crop gradients; combined with a refined patch-size ensemble (PE+), this strengthens transferable directions. Together these modules form M-Attack-V2, a simple, modular enhancement over M-Attack that substantially improves transfer-based black-box attacks on frontier LVLMs: boosting success rates on Claude-4.0 from 8% to 30%, Gemini-2.5-Pro from 83% to 97%, and GPT-5 from 98% to 100%, outperforming prior black-box LVLM attacks. Code and data are publicly available at: https://github.com/vila-lab/M-Attack-V2.

With the rapid advancement of Large Language Models (LLMs), the safety of LLMs has been a critical concern requiring precise assessment. Current benchmarks primarily concentrate on single-turn dialogues or a single jailbreak attack method to assess the safety. Additionally, these benchmarks have not taken into account the LLM's capability of identifying and handling unsafe information in detail. To address these issues, we propose a fine-grained benchmark SafeDialBench for evaluating the safety of LLMs across various jailbreak attacks in multi-turn dialogues. Specifically, we design a two-tier hierarchical safety taxonomy that considers 6 safety dimensions and generates more than 4000 multi-turn dialogues in both Chinese and English under 22 dialogue scenarios. We employ 7 jailbreak attack strategies, such as reference attack and purpose reverse, to enhance the dataset quality for dialogue generation. Notably, we construct an innovative assessment framework of LLMs, measuring capabilities in detecting, and handling unsafe information and maintaining consistency when facing jailbreak attacks. Experimental results across 17 LLMs reveal that Yi-34B-Chat and GLM4-9B-Chat demonstrate superior safety performance, while Llama3.1-8B-Instruct and o3-mini exhibit safety vulnerabilities.

We present FACESEC, a framework for fine-grained robustness evaluation of face recognition systems. FACESEC evaluation is performed along four dimensions of adversarial modeling: the nature of perturbation (e.g., pixel-level or face accessories), the attacker's system knowledge (about training data and learning architecture), goals (dodging or impersonation), and capability (tailored to individual inputs or across sets of these). We use FACESEC to study five face recognition systems in both closed-set and open-set settings, and to evaluate the state-of-the-art approach for defending against physically realizable attacks on these. We find that accurate knowledge of neural architecture is significantly more important than knowledge of the training data in black-box attacks. Moreover, we observe that open-set face recognition systems are more vulnerable than closed-set systems under different types of attacks. The efficacy of attacks for other threat model variations, however, appears highly dependent on both the nature of perturbation and the neural network architecture. For example, attacks that involve adversarial face masks are usually more potent, even against adversarially trained models, and the ArcFace architecture tends to be more robust than the others.

The truthfulness of existing explanation methods in authentically elucidating the underlying model's decision-making process has been questioned. Existing methods have deviated from faithfully representing the model, thus susceptible to adversarial attacks. To address this, we propose a novel eXplainable AI (XAI) method called SRD (Sharing Ratio Decomposition), which sincerely reflects the model's inference process, resulting in significantly enhanced robustness in our explanations. Different from the conventional emphasis on the neuronal level, we adopt a vector perspective to consider the intricate nonlinear interactions between filters. We also introduce an interesting observation termed Activation-Pattern-Only Prediction (APOP), letting us emphasize the importance of inactive neurons and redefine relevance encapsulating all relevant information including both active and inactive neurons. Our method, SRD, allows for the recursive decomposition of a Pointwise Feature Vector (PFV), providing a high-resolution Effective Receptive Field (ERF) at any layer.

Membership inference attacks aim to detect if a particular data point was used in training a model. We design a novel statistical test to perform robust membership inference attacks (RMIA) with low computational overhead. We achieve this by a fine-grained modeling of the null hypothesis in our likelihood ratio tests, and effectively leveraging both reference models and reference population data samples. RMIA has superior test power compared with prior methods, throughout the TPR-FPR curve (even at extremely low FPR, as low as 0). Under computational constraints, where only a limited number of pre-trained reference models (as few as 1) are available, and also when we vary other elements of the attack (e.g., data distribution), our method performs exceptionally well, unlike prior attacks that approach random guessing. RMIA lays the groundwork for practical yet accurate data privacy risk assessment in machine learning.

The Model Context Protocol (MCP) plays a crucial role in extending the capabilities of Large Language Models (LLMs) by enabling integration with external tools and data sources. However, the standard MCP specification presents significant security vulnerabilities, notably Tool Poisoning and Rug Pull attacks. This paper introduces the Enhanced Tool Definition Interface (ETDI), a security extension designed to fortify MCP. ETDI incorporates cryptographic identity verification, immutable versioned tool definitions, and explicit permission management, often leveraging OAuth 2.0. We further propose extending MCP with fine-grained, policy-based access control, where tool capabilities are dynamically evaluated against explicit policies using a dedicated policy engine, considering runtime context beyond static OAuth scopes. This layered approach aims to establish a more secure, trustworthy, and controllable ecosystem for AI applications interacting with LLMs and external tools.

Federated large language models (FedLLMs) enable cross-silo collaborative training among institutions while preserving data locality, making them appealing for privacy-sensitive domains such as law, finance, and healthcare. However, the memorization behavior of LLMs can lead to privacy risks that may cause cross-client data leakage. In this work, we study the threat of cross-client data extraction, where a semi-honest participant attempts to recover personally identifiable information (PII) memorized from other clients' data. We propose three simple yet effective extraction strategies that leverage contextual prefixes from the attacker's local data, including frequency-based prefix sampling and local fine-tuning to amplify memorization. To evaluate these attacks, we construct a Chinese legal-domain dataset with fine-grained PII annotations consistent with CPIS, GDPR, and CCPA standards, and assess extraction performance using two metrics: coverage and efficiency. Experimental results show that our methods can recover up to 56.6% of victim-exclusive PII, where names, addresses, and birthdays are particularly vulnerable. These findings highlight concrete privacy risks in FedLLMs and establish a benchmark and evaluation framework for future research on privacy-preserving federated learning. Code and data are available at https://github.com/SMILELab-FL/FedPII.

Machine-learning phishing webpage detectors (ML-PWD) have been shown to suffer from adversarial manipulations of the HTML code of the input webpage. Nevertheless, the attacks recently proposed have demonstrated limited effectiveness due to their lack of optimizing the usage of the adopted manipulations, and they focus solely on specific elements of the HTML code. In this work, we overcome these limitations by first designing a novel set of fine-grained manipulations which allow to modify the HTML code of the input phishing webpage without compromising its maliciousness and visual appearance, i.e., the manipulations are functionality- and rendering-preserving by design. We then select which manipulations should be applied to bypass the target detector by a query-efficient black-box optimization algorithm. Our experiments show that our attacks are able to raze to the ground the performance of current state-of-the-art ML-PWD using just 30 queries, thus overcoming the weaker attacks developed in previous work, and enabling a much fairer robustness evaluation of ML-PWD.

Jailbreaking is an essential adversarial technique for red-teaming these models to uncover and patch security flaws. However, existing jailbreak methods face significant drawbacks. Token-level jailbreak attacks often produce incoherent or unreadable inputs and exhibit poor transferability, while prompt-level attacks lack scalability and rely heavily on manual effort and human ingenuity. We propose a concise and effective two-stage framework that combines the advantages of these approaches. The first stage performs a scenario-based generation of context and rephrases the original malicious query to obscure its harmful intent. The second stage then utilizes information from the model's hidden states to guide fine-grained edits, effectively steering the model's internal representation of the input from a malicious toward a benign one. Extensive experiments demonstrate that this method achieves state-of-the-art Attack Success Rate, with gains of up to 37.74% over the strongest baseline, and exhibits excellent transferability to black-box models. Our analysis further demonstrates that AGILE maintains substantial effectiveness against prominent defense mechanisms, highlighting the limitations of current safeguards and providing valuable insights for future defense development. Our code is available at https://github.com/yunsaijc/AGILE.

Multimodal large language models (MLLMs) remain vulnerable to transferable adversarial examples. While existing methods typically achieve targeted attacks by aligning global features-such as CLIP's [CLS] token-between adversarial and target samples, they often overlook the rich local information encoded in patch tokens. This leads to suboptimal alignment and limited transferability, particularly for closed-source models. To address this limitation, we propose a targeted transferable adversarial attack method based on feature optimal alignment, called FOA-Attack, to improve adversarial transfer capability. Specifically, at the global level, we introduce a global feature loss based on cosine similarity to align the coarse-grained features of adversarial samples with those of target samples. At the local level, given the rich local representations within Transformers, we leverage clustering techniques to extract compact local patterns to alleviate redundant local features. We then formulate local feature alignment between adversarial and target samples as an optimal transport (OT) problem and propose a local clustering optimal transport loss to refine fine-grained feature alignment. Additionally, we propose a dynamic ensemble model weighting strategy to adaptively balance the influence of multiple models during adversarial example generation, thereby further improving transferability. Extensive experiments across various models demonstrate the superiority of the proposed method, outperforming state-of-the-art methods, especially in transferring to closed-source MLLMs. The code is released at https://github.com/jiaxiaojunQAQ/FOA-Attack.

Multi-targeted adversarial attacks aim to mislead classifiers toward specific target classes using a single perturbation generator with a conditional input specifying the desired target class. Existing methods face two key limitations: (1) a single generator supports only a limited number of predefined target classes, and (2) it requires access to the victim model's training data to learn target class semantics. This dependency raises data leakage concerns in practical black-box scenarios where the training data is typically private. To address these limitations, we propose a novel Cross-Domain Multi-Targeted Attack (CD-MTA) that can generate perturbations toward arbitrary target classes, even those that do not exist in the attacker's training data. CD-MTA is trained on a single public dataset but can perform targeted attacks on black-box models trained on different datasets with disjoint and unknown class sets. Our method requires only a single example image that visually represents the desired target class, without relying its label, class distribution or pretrained embeddings. We achieve this through a Feature Injection Module (FIM) and class-agnostic objectives which guide the generator to extract transferable, fine-grained features from the target image without inferring class semantics. Experiments on ImageNet and seven additional datasets show that CD-MTA outperforms existing multi-targeted attack methods on unseen target classes in black-box and cross-domain scenarios. The code is available at https://github.com/tgoncalv/CD-MTA.

Large Language Models (LLMs) are increasingly used in a variety of applications, but concerns around membership inference have grown in parallel. Previous efforts focus on black-to-grey-box models, thus neglecting the potential benefit from internal LLM information. To address this, we propose the use of Linear Probes (LPs) as a method to detect Membership Inference Attacks (MIAs) by examining internal activations of LLMs. Our approach, dubbed LUMIA, applies LPs layer-by-layer to get fine-grained data on the model inner workings. We test this method across several model architectures, sizes and datasets, including unimodal and multimodal tasks. In unimodal MIA, LUMIA achieves an average gain of 15.71 % in Area Under the Curve (AUC) over previous techniques. Remarkably, LUMIA reaches AUC>60% in 65.33% of cases -- an increment of 46.80% against the state of the art. Furthermore, our approach reveals key insights, such as the model layers where MIAs are most detectable. In multimodal models, LPs indicate that visual inputs can significantly contribute to detect MIAs -- AUC>60% is reached in 85.90% of experiments.

Emerging AR-LLM-based Social Engineering attack (e.g., SEAR) is at the edge of posing great threats to real-world social life. In such AR-LLM-SE attack, the attacker can leverage AR (Augmented Reality) glass to capture the image and vocal information of the target, using the LLM to identify the target and generate the social profile, using the LLM agents to apply social engineering strategies for conversation suggestion to win the target trust and perform phishing afterwards. Current defensive approaches, such as role-based access control or data flow tracking, are not directly applicable to the convergent AR-LLM ecosystem (considering embedded AR device and opaque LLM inference), leaving an emerging and potent social engineering threat that existing privacy paradigms are ill-equipped to address. This necessitates a shift beyond solely human-centric measures like legislation and user education toward enforceable vendor policies and platform-level restrictions. Realizing this vision, however, faces significant technical challenges: securing resource-constrained AR-embedded devices, implementing fine-grained access control within opaque LLM inferences, and governing adaptive interactive agents. To address these challenges, we present UNSEEN, a coordinated cross-stack defense that combines an AR ACL (Access Control Layer) for identity-gated sensing, F-RMU-based LLM unlearning for sensitive profile suppression, and runtime agent guardrails for adaptive interaction control. We evaluate UNSEEN in an IRB-approved user study with 60 participants and a dataset of 360 annotated conversations across realistic social scenarios.

Multiple prompt injection attacks have been proposed against web agents. At the same time, various methods have been developed to detect general prompt injection attacks, but none have been systematically evaluated for web agents. In this work, we bridge this gap by presenting the first comprehensive benchmark study on detecting prompt injection attacks targeting web agents. We begin by introducing a fine-grained categorization of such attacks based on the threat model. We then construct datasets containing both malicious and benign samples: malicious text segments generated by different attacks, benign text segments from four categories, malicious images produced by attacks, and benign images from two categories. Next, we systematize both text-based and image-based detection methods. Finally, we evaluate their performance across multiple scenarios. Our key findings show that while some detectors can identify attacks that rely on explicit textual instructions or visible image perturbations with moderate to high accuracy, they largely fail against attacks that omit explicit instructions or employ imperceptible perturbations. Our datasets and code are released at: https://github.com/Norrrrrrr-lyn/WAInjectBench.

Generative models are rapidly gaining popularity and being integrated into everyday applications, raising concerns over their safe use as various vulnerabilities are exposed. In light of this, the field of red teaming is undergoing fast-paced growth, highlighting the need for a comprehensive survey covering the entire pipeline and addressing emerging topics. Our extensive survey, which examines over 120 papers, introduces a taxonomy of fine-grained attack strategies grounded in the inherent capabilities of language models. Additionally, we have developed the "searcher" framework to unify various automatic red teaming approaches. Moreover, our survey covers novel areas including multimodal attacks and defenses, risks around LLM-based agents, overkill of harmless queries, and the balance between harmlessness and helpfulness.

Vision-Language-Action (VLA) models map multimodal perception and language instructions to executable robot actions, making them particularly vulnerable to behavioral backdoor manipulation: a hidden trigger introduced during training can induce unintended physical actions while nominal task performance remains intact. Prior work on VLA backdoors primarily studies untargeted attacks or task-level hijacking, leaving fine-grained control over individual actions largely unexplored. In this work, we present DropVLA, an action-level backdoor attack that forces a reusable action primitive (e.g., open_gripper) to execute at attacker-chosen decision points under a realistic pipeline-black-box setting with limited data-poisoning access, using a window-consistent relabeling scheme for chunked fine-tuning. On OpenVLA-7B evaluated with LIBERO, vision-only poisoning achieves 98.67%-99.83% attack success rate (ASR) with only 0.31% poisoned episodes while preserving 98.50%-99.17% clean-task retention, and successfully triggers the targeted action within 25 control steps at 500 Hz (0.05 s). Text-only triggers are unstable at low poisoning budgets, and combining text with vision provides no consistent ASR improvement over vision-only attacks. The backdoor remains robust to moderate trigger variations and transfers across evaluation suites (96.27%, 99.09%), whereas text-only largely fails (0.72%). We further validate physical-world feasibility on a 7-DoF Franka arm with pi0-fast, demonstrating non-trivial attack efficacy under camera-relative motion that induces image-plane trigger drift. These results reveal that VLA models can be covertly steered at the granularity of safety-critical actions with minimal poisoning and without observable degradation of nominal performance.

Prompt injection attacks pose a critical threat to large language models (LLMs), enabling goal hijacking and data leakage. Prompt guard models, though effective in defense, suffer from over-defense -- falsely flagging benign inputs as malicious due to trigger word bias. To address this issue, we introduce NotInject, an evaluation dataset that systematically measures over-defense across various prompt guard models. NotInject contains 339 benign samples enriched with trigger words common in prompt injection attacks, enabling fine-grained evaluation. Our results show that state-of-the-art models suffer from over-defense issues, with accuracy dropping close to random guessing levels (60%). To mitigate this, we propose InjecGuard, a novel prompt guard model that incorporates a new training strategy, Mitigating Over-defense for Free (MOF), which significantly reduces the bias on trigger words. InjecGuard demonstrates state-of-the-art performance on diverse benchmarks including NotInject, surpassing the existing best model by 30.8%, offering a robust and open-source solution for detecting prompt injection attacks. The code and datasets are released at https://github.com/SaFoLab-WISC/InjecGuard.

Current multimodal toxicity benchmarks typically use a single binary hatefulness label. This coarse approach conflates two fundamentally different characteristics of expression: tone and content. Drawing on communication science theory, we introduce a fine-grained annotation scheme that distinguishes two separable dimensions: incivility (rude or dismissive tone) and intolerance (content that attacks pluralism and targets groups or identities) and apply it to 2,030 memes from the Hateful Memes dataset. We evaluate different vision-language models under coarse-label training, transfer learning across label schemes and a joint learning approach that combines the coarse hatefulness label with our fine-grained annotations. Our results show that fine-grained annotations complement existing coarse labels and, when used jointly, improve overall model performance. Moreover, models trained with the fine-grained scheme exhibit more balanced moderation-relevant error profiles and are less prone to under-detection of harmful content than models trained on hatefulness labels alone (FNR-FPR, the difference between false negative and false positive rates: 0.74 to 0.42 for LLaVA-1.6-Mistral-7B; 0.54 to 0.28 for Qwen2.5-VL-7B). This work contributes to data-centric approaches in content moderation by improving the reliability and accuracy of moderation systems through enhanced data quality. Overall, combining both coarse and fine-grained labels provides a practical route to more reliable multimodal moderation.

Personal digital data is a critical asset, and governments worldwide have enforced laws and regulations to protect data privacy. Data users have been endowed with the right to be forgotten of their data. In the course of machine learning (ML), the forgotten right requires a model provider to delete user data and its subsequent impact on ML models upon user requests. Machine unlearning emerges to address this, which has garnered ever-increasing attention from both industry and academia. While the area has developed rapidly, there is a lack of comprehensive surveys to capture the latest advancements. Recognizing this shortage, we conduct an extensive exploration to map the landscape of machine unlearning including the (fine-grained) taxonomy of unlearning algorithms under centralized and distributed settings, debate on approximate unlearning, verification and evaluation metrics, challenges and solutions for unlearning under different applications, as well as attacks targeting machine unlearning. The survey concludes by outlining potential directions for future research, hoping to serve as a guide for interested scholars.

Machine unlearning for large language models (LLMs) aims to remove undesired data, knowledge, and behaviors (e.g., for safety, privacy, or copyright) while preserving useful model capabilities. Despite rapid progress over the past two years, research in LLM unlearning remains fragmented, with limited clarity on what constitutes effective unlearning and how it should be rigorously evaluated. In this work, we present a principled taxonomy of twelve recent stateful unlearning methods, grouped into three methodological families: divergence-driven optimization, representation misalignment, and rejection-based targeted unlearning. Building on this taxonomy, we revisit the evaluation of unlearning effectiveness (UE), utility retention (UT), and robustness (Rob), focusing on the WMDP benchmark. Our analysis shows that current evaluations, dominated by multiple-choice question (MCQ) accuracy, offer only a narrow perspective, often overstating success while overlooking the model's actual generation behavior. To address this gap, we introduce open question-answering (Open-QA) metrics that better capture generative performance and reveal the inherent UE-UT tradeoff across method families. Furthermore, we demonstrate that robustness requires finer-grained analysis: for example, vulnerabilities differ substantially between in-domain relearning and out-of-domain fine-tuning, even though both fall under model-level attacks. Through this study, we hope to deliver a full-stack revisit of LLM unlearning and actionable guidance for designing and evaluating future methods.

Hateful memes are a growing form of multimodal online harm, where hostile intent is often conveyed through the joint interpretation of images, text, cultural references, and implicit targets. While hateful meme detection has advanced in high-resource languages, Arabic remains underexplored, with existing meme resources focusing mainly on propaganda or coarse harmful-content labels. We introduce AHA-Memes (Arabic HAteful Memes), which is, to our knowledge, the first large-scale Arabic hateful meme benchmark with fine-grained, multi-label annotations. The dataset includes 5K manually annotated memes using a taxonomy that captures hate types, i.e., attack strategies. We further provide ~66K silver-labeled memes to support future studies. We benchmark text-only, image-only, and late-fusion multimodal models, as well as few-shot in-context learning (ICL) and open- and closed-weight Vision-Language Models (VLMs) under zero-shot and fine-tuning settings. Our results establish strong baselines and highlight key challenges in culturally grounded Arabic hateful meme detection. We release the dataset, annotation guidelines, and evaluation scripts to support future research. WARNING: This paper contains examples that may be disturbing to readers.

Large language model (LLM) agents increasingly operate through long-horizon trajectories involving user instructions, tool use, external observations, and memory. Existing benchmarks primarily evaluate behavioral outcomes but provide limited support for fine-grained attribution analysis. We introduce trajectory attribution and develop a benchmark and annotation framework for this task. The benchmark organizes heterogeneous trajectories under a unified component schema and provides annotations of the primary attribution component, together with attack and execution chains where applicable. Instantiating the benchmark with trajectories from AgentDojo and the Stage and Canary settings of Agent3Sigma yields more than 1,300 annotated trajectories covering task-aligned actions, unsafe actions, and safety refusals. The benchmark defines two evaluation tasks, primary attribution localization and attribution-chain recovery, and provides reference baselines based on incremental trajectory contribution and component-level leave-one-out perturbation. It captures diverse attribution settings, including local and long-range attribution as well as structured attribution chains. Reference baseline results exhibit substantial performance differences across these settings, providing an initial characterization of the benchmark's attribution challenges. Beyond this initial instantiation, we release a reusable annotation skill that enables trajectories generated by new agent models to be standardized, annotated, and evaluated under the same framework. Project resources and future releases are available at https://github.com/chenjing-2024/agent-trajectory-attribution.

The generation of transferable adversarial perturbations typically involves training a generator to maximize embedding separation between clean and adversarial images at a single mid-layer of a source model. In this work, we build on this approach and introduce Neuron Attack for Transferability (NAT), a method designed to target specific neuron within the embedding. Our approach is motivated by the observation that previous layer-level optimizations often disproportionately focus on a few neurons representing similar concepts, leaving other neurons within the attacked layer minimally affected. NAT shifts the focus from embedding-level separation to a more fundamental, neuron-specific approach. We find that targeting individual neurons effectively disrupts the core units of the neural network, providing a common basis for transferability across different models. Through extensive experiments on 41 diverse ImageNet models and 9 fine-grained models, NAT achieves fooling rates that surpass existing baselines by over 14\% in cross-model and 4\% in cross-domain settings. Furthermore, by leveraging the complementary attacking capabilities of the trained generators, we achieve impressive fooling rates within just 10 queries. Our code is available at: https://krishnakanthnakka.github.io/NAT/

Although multimodal large language models (MLLMs) are increasingly deployed in real-world applications, their instruction-following behavior leaves them vulnerable to prompt injection attacks. Existing prompt injection methods predominantly rely on textual prompts or perceptible visual prompts that are observable by human users. In this work, we study imperceptible visual prompt injection against powerful closed-source MLLMs, where adversarial instructions are embedded in the visual modality. Our method adaptively embeds the malicious prompt into the input image via a bounded text overlay to provide semantic guidance. Meanwhile, the imperceptible visual perturbation is iteratively optimized to align the feature representation of the attacked image with those of the malicious visual and textual targets at both coarse- and fine-grained levels. Specifically, the visual target is instantiated as a text-rendered image and progressively refined during optimization to more faithfully represent the desired semantics and improve transferability. Extensive experiments on two multimodal understanding tasks across multiple closed-source MLLMs demonstrate the superior performance of our approach compared to existing methods.

Recent advances in multi-modal large language models (MLLMs) have enabled unified perception-reasoning capabilities, yet these systems remain highly vulnerable to jailbreak attacks that bypass safety alignment and induce harmful behaviors. Existing benchmarks such as JailBreakV-28K, MM-SafetyBench, and HADES provide valuable insights into multi-modal vulnerabilities, but they typically focus on limited attack scenarios, lack standardized defense evaluation, and offer no unified, reproducible toolbox. To address these gaps, we introduce OmniSafeBench-MM, which is a comprehensive toolbox for multi-modal jailbreak attack-defense evaluation. OmniSafeBench-MM integrates 13 representative attack methods, 15 defense strategies, and a diverse dataset spanning 9 major risk domains and 50 fine-grained categories, structured across consultative, imperative, and declarative inquiry types to reflect realistic user intentions. Beyond data coverage, it establishes a three-dimensional evaluation protocol measuring (1) harmfulness, distinguished by a granular, multi-level scale ranging from low-impact individual harm to catastrophic societal threats, (2) intent alignment between responses and queries, and (3) response detail level, enabling nuanced safety-utility analysis. We conduct extensive experiments on 10 open-source and 8 closed-source MLLMs to reveal their vulnerability to multi-modal jailbreak. By unifying data, methodology, and evaluation into an open-source, reproducible platform, OmniSafeBench-MM provides a standardized foundation for future research. The code is released at https://github.com/jiaxiaojunQAQ/OmniSafeBench-MM.

Existing agent-safety evaluation has focused mainly on externally induced risks. Yet agents may still enter unsafe trajectories under benign conditions. We study this complementary but underexplored setting through the lens of intrinsic risk, where intrinsic failures remain latent, propagate across long-horizon execution, and eventually lead to high-consequence outcomes. To evaluate this setting, we introduce non-attack intrinsic risk auditing and present HINTBench, a benchmark of 629 agent trajectories (523 risky, 106 safe; 33 steps on average) supporting three tasks: risk detection, risk-step localization, and intrinsic failure-type identification. Its annotations are organized under a unified five-constraint taxonomy. Experiments reveal a substantial capability gap: strong LLMs perform well on trajectory-level risk detection, but their performance drops to below 35 Strict-F1 on risk-step localization, while fine-grained failure diagnosis proves even harder. Existing guard models transfer poorly to this setting. These findings establish intrinsic risk auditing as an open challenge for agent safety.

Recent studies have shown that large language models (LLMs) can infer private user attributes (e.g., age, location, gender) from user-generated text shared online, enabling rapid and large-scale privacy breaches. Existing anonymization-based defenses are coarse-grained, lacking word-level precision in anonymizing privacy-leaking elements. Moreover, they are inherently limited as altering user text to hide sensitive cues still allows attribute inference to occur through models' reasoning capabilities. To address these limitations, we propose a unified defense framework that combines fine-grained anonymization (TRACE) with inference-preventing optimization (RPS). TRACE leverages attention mechanisms and inference chain generation to identify and anonymize privacy-leaking textual elements, while RPS employs a lightweight two-stage optimization strategy to induce model rejection behaviors, thereby preventing attribute inference. Evaluations across diverse LLMs show that TRACE-RPS reduces attribute inference accuracy from around 50\% to below 5\% on open-source models. In addition, our approach offers strong cross-model generalization, prompt-variation robustness, and utility-privacy tradeoffs. Our code is available at https://github.com/Jasper-Yan/TRACE-RPS.

Using risky text prompts, such as pornography and violent prompts, to test the safety of text-to-image (T2I) models is a critical task. However, existing risky prompt datasets are limited in three key areas: 1) limited risky categories, 2) coarse-grained annotation, and 3) low effectiveness. To address these limitations, we introduce T2I-RiskyPrompt, a comprehensive benchmark designed for evaluating safety-related tasks in T2I models. Specifically, we first develop a hierarchical risk taxonomy, which consists of 6 primary categories and 14 fine-grained subcategories. Building upon this taxonomy, we construct a pipeline to collect and annotate risky prompts. Finally, we obtain 6,432 effective risky prompts, where each prompt is annotated with both hierarchical category labels and detailed risk reasons. Moreover, to facilitate the evaluation, we propose a reason-driven risky image detection method that explicitly aligns the MLLM with safety annotations. Based on T2I-RiskyPrompt, we conduct a comprehensive evaluation of eight T2I models, nine defense methods, five safety filters, and five attack strategies, offering nine key insights into the strengths and limitations of T2I model safety. Finally, we discuss potential applications of T2I-RiskyPrompt across various research fields. The dataset and code are provided in https://github.com/datar001/T2I-RiskyPrompt.

Large language models can hide hidden behaviors that activate only under narrow conditions, such as backdoor triggers, sleeper-agent deployment cues, sandbagging, or topic-conditioned censorship. Such behaviors are difficult to detect without prior knowledge what to look for. We present activation-matched finetuning, an unsupervised detection method that assumes no knowledge of the trigger or the target behavior. Given a suspect model and a publicly available anchor, we finetune the anchor to reproduce the suspect's activations on a small benign corpus, and score each evaluation prompt by the residual between the two models. Since no benign corpus covers the sparse trigger region, the reference learns the benign computation but not the hidden behavior. Therefore, trigger prompts -- and, crucially, their semantic neighbors -- incur a large residual that signal the presence of unusual behavior to the defender. Testing our method across third-party models and custom models, activation-matched finetuning surfaces hidden behavior reliably. Furthermore, we empirically consider a natural defense-aware attack and showcase that it fails to suppress our detection method without sacrificing the behavior itself.

We present Virtual Prompt Injection (VPI) for instruction-tuned Large Language Models (LLMs). VPI allows an attacker-specified virtual prompt to steer the model behavior under specific trigger scenario without any explicit injection in model input. For instance, if an LLM is compromised with the virtual prompt "Describe Joe Biden negatively." for Joe Biden-related instructions, then any service deploying this model will propagate biased views when handling user queries related to Joe Biden. VPI is especially harmful for two primary reasons. Firstly, the attacker can take fine-grained control over LLM behaviors by defining various virtual prompts, exploiting LLMs' proficiency in following instructions. Secondly, this control is achieved without any interaction from the attacker while the model is in service, leading to persistent attack. To demonstrate the threat, we propose a simple method for performing VPI by poisoning the model's instruction tuning data. We find that our proposed method is highly effective in steering the LLM with VPI. For example, by injecting only 52 poisoned examples (0.1% of the training data size) into the instruction tuning data, the percentage of negative responses given by the trained model on Joe Biden-related queries change from 0% to 40%. We thus highlight the necessity of ensuring the integrity of the instruction-tuning data as little poisoned data can cause stealthy and persistent harm to the deployed model. We further explore the possible defenses and identify data filtering as an effective way to defend against the poisoning attacks. Our project page is available at https://poison-llm.github.io.

In recent years, safety risks associated with large language models have become increasingly prominent, highlighting the urgent need to mitigate the generation of toxic and harmful content. The mainstream paradigm for LLM safety alignment typically adopts a collaborative framework involving three roles: an attacker for adversarial prompt generation, a defender for safety defense, and an evaluator for response assessment. In this paper, we propose a closed-loop reinforcement learning framework called TriPlay-RL that enables iterative and co-improving collaboration among three roles with near-zero manual annotation. Experimental results show that the attacker preserves high output diversity while achieving a 20%-50% improvement in adversarial effectiveness; the defender attains 10%-30% gains in safety performance without degrading general reasoning capability; and the evaluator continuously refines its fine-grained judgment ability through iterations, accurately distinguishing unsafe responses, simple refusals, and useful guidance. Overall, our framework establishes an efficient and scalable paradigm for LLM safety alignment, enabling continuous co-evolution within a unified learning loop.

We present MultiBreak, a scalable and diverse multi-turn jailbreak benchmark to evaluate large language model (LLM) safety. Multi-turn jailbreaks mimic natural conversational settings, making them easier to bypass safety-aligned LLM than single-turn jailbreaks. Existing multi-turn benchmarks are limited in size or rely heavily on templates, which restrict their diversity. To address this gap, we unify a wide range of harmful jailbreak intents, and introduce an active learning pipeline for expanding high-quality multi-turn adversarial prompts, where a generator is iteratively fine-tuned to produce stronger attack candidates, guided by uncertainty-based refinement. Our MultiBreak includes 10,389 multi-turn adversarial prompts, spans 2,665 distinct harmful intents, and covers the most diverse set of topics to date. Empirical evaluation shows that our benchmark achieves up to a 54.0 and 34.6 higher attack success rate (ASR)} than the second-best dataset on DeepSeek-R1-7B and GPT-4.1-mini, respectively. More importantly, safety evaluations suggest that diverse attack categories uncover fine-grained LLM vulnerabilities}, and categories that appear benign under single-turn can exhibit substantially higher adversarial effectiveness in multi-turn scenarios. These findings highlight persistent vulnerabilities of LLMs under realistic adversarial settings and establish MultiBreak as a scalable resource for advancing LLM safety.

This technical report accompanies the LLM robustness leaderboard published by PRISM Eval for the Paris AI Action Summit. We introduce PRISM Eval Behavior Elicitation Tool (BET), an AI system performing automated red-teaming through Dynamic Adversarial Optimization that achieves 100% Attack Success Rate (ASR) against 37 of 41 state-of-the-art LLMs. Beyond binary success metrics, we propose a fine-grained robustness metric estimating the average number of attempts required to elicit harmful behaviors, revealing that attack difficulty varies by over 300-fold across models despite universal vulnerability. We introduce primitive-level vulnerability analysis to identify which jailbreaking techniques are most effective for specific hazard categories. Our collaborative evaluation with trusted third parties from the AI Safety Network demonstrates practical pathways for distributed robustness assessment across the community.

Vision-language models (VLMs) are increasingly deployed in consumer, medical, financial, and enterprise applications. This broad deployment expands the safety surface: risks can arise from multimodal question answering, assistant responses, and cross-modal composition, while moderation policies may vary across products, regions, and deployment stages. Most existing guardrails either rely on fixed taxonomies or target only a narrow set of interaction settings, which limits their adaptability when safety rules change at deployment time. We present SingGuard, a policy-adaptive multimodal guardrail model family for safety assessment in multimodal conversations. SingGuard treats the active policy as a runtime input: given natural-language rules, it checks the target content against the active policy rule by rule and predicts both the safety label and the triggered rule. To balance efficiency and interpretability, SingGuard supports fast, hybrid, and slow inference regimes along a fast-to-slow reasoning spectrum, ranging from direct safety judgments to policy-grounded deliberation. We further optimize this behavior with fast--slow decoupled reinforcement learning. We also introduce SingGuard-Bench, a multimodal guardrail benchmark with 56{,}340 examples spanning 80+ fine-grained risk types across multimodal QA, adversarial attack, and dynamic-rule evaluation settings, including cross-modal joint-risk cases where each modality is harmless in isolation but their composition implies unsafe intent. Across six benchmark families (35 datasets), SingGuard achieves state-of-the-art average F1 in every family. Dynamic-rule evaluation further shows improved policy-following accuracy from 0.6465 to 0.7415 under runtime policy shifts. Our code is available at https://github.com/inclusionAI/Sing-Guard.

Hand-crafted image quality metrics, such as PSNR and SSIM, are commonly used to evaluate model privacy risk under reconstruction attacks. Under these metrics, reconstructed images that are determined to resemble the original one generally indicate more privacy leakage. Images determined as overall dissimilar, on the other hand, indicate higher robustness against attack. However, there is no guarantee that these metrics well reflect human opinions, which, as a judgement for model privacy leakage, are more trustworthy. In this paper, we comprehensively study the faithfulness of these hand-crafted metrics to human perception of privacy information from the reconstructed images. On 5 datasets ranging from natural images, faces, to fine-grained classes, we use 4 existing attack methods to reconstruct images from many different classification models and, for each reconstructed image, we ask multiple human annotators to assess whether this image is recognizable. Our studies reveal that the hand-crafted metrics only have a weak correlation with the human evaluation of privacy leakage and that even these metrics themselves often contradict each other. These observations suggest risks of current metrics in the community. To address this potential risk, we propose a learning-based measure called SemSim to evaluate the Semantic Similarity between the original and reconstructed images. SemSim is trained with a standard triplet loss, using an original image as an anchor, one of its recognizable reconstructed images as a positive sample, and an unrecognizable one as a negative. By training on human annotations, SemSim exhibits a greater reflection of privacy leakage on the semantic level. We show that SemSim has a significantly higher correlation with human judgment compared with existing metrics. Moreover, this strong correlation generalizes to unseen datasets, models and attack methods.

We describe a vulnerability in language models (LMs) trained with user feedback, whereby a single user can persistently alter LM knowledge and behavior given only the ability to provide prompts and upvote / downvote feedback on LM outputs. To implement the attack, the attacker prompts the LM to stochastically output either a "poisoned" or benign response, then upvotes the poisoned response or downvotes the benign one. When feedback signals are used in a subsequent preference tuning behavior, LMs exhibit increased probability of producing poisoned responses even in contexts without malicious prompts. We show that this attack can be used to (1) insert factual knowledge the model did not previously possess, (2) modify code generation patterns in ways that introduce exploitable security flaws, and (3) inject fake financial news. Our finding both identifies a new qualitative feature of language model preference tuning (showing that it even highly restricted forms of preference data can be used to exert fine-grained control over behavior), and a new attack mechanism for LMs trained with user feedback (extending work on pretraining-time data poisoning and deployment-time prompt injection).

When building Large Language Models (LLMs), it is paramount to bear safety in mind and protect them with guardrails. Indeed, LLMs should never generate content promoting or normalizing harmful, illegal, or unethical behavior that may contribute to harm to individuals or society. This principle applies to both normal and adversarial use. In response, we introduce ALERT, a large-scale benchmark to assess safety based on a novel fine-grained risk taxonomy. It is designed to evaluate the safety of LLMs through red teaming methodologies and consists of more than 45k instructions categorized using our novel taxonomy. By subjecting LLMs to adversarial testing scenarios, ALERT aims to identify vulnerabilities, inform improvements, and enhance the overall safety of the language models. Furthermore, the fine-grained taxonomy enables researchers to perform an in-depth evaluation that also helps one to assess the alignment with various policies. In our experiments, we extensively evaluate 10 popular open- and closed-source LLMs and demonstrate that many of them still struggle to attain reasonable levels of safety.

Current open-weight large language models (LLMs) are prone to malicious finetuning attacks, which could compromise the safety alignment of LLMs with only a few steps of supervised finetuning (SFT) on poisoned datasets. Existing alignment-stage defenses are primarily designed to defend against attacks that use parameter-efficient finetuning methods. However, they fail to defend against stronger attacks that use full-parameter finetuning. In this paper, we propose Patcher, a method inspired by adversarial training and bi-level optimization, to combat such attacks. Patcher strengthens the simulated attack by scaling up the optimization steps in the adversarial loop, thus forcing the defender to find model parameters that are insensitive to stronger attacks. Furthermore, we propose an efficient parallel algorithm to implement Patcher, decreasing the wall-clock time of training while preserving Patcher's performance. Extensive experiments show that Patcher substantially improves the model's robustness compared to vanilla SFT alignment, and transfers to diverse attack scenarios and model sizes. Code is available at https://github.com/haomingwen/patcher.

Language model attacks typically assume one of two extreme threat models: full white-box access to model weights, or black-box access limited to a text generation API. However, real-world APIs are often more flexible than just text generation: these APIs expose "gray-box" access leading to new threat vectors. To explore this, we red-team three new functionalities exposed in the GPT-4 APIs: fine-tuning, function calling and knowledge retrieval. We find that fine-tuning a model on as few as 15 harmful examples or 100 benign examples can remove core safeguards from GPT-4, enabling a range of harmful outputs. Furthermore, we find that GPT-4 Assistants readily divulge the function call schema and can be made to execute arbitrary function calls. Finally, we find that knowledge retrieval can be hijacked by injecting instructions into retrieval documents. These vulnerabilities highlight that any additions to the functionality exposed by an API can create new vulnerabilities.

LLM developers have imposed technical interventions to prevent fine-tuning misuse attacks, attacks where adversaries evade safeguards by fine-tuning the model using a public API. Previous work has established several successful attacks against specific fine-tuning API defences. In this work, we show that defences of fine-tuning APIs that seek to detect individual harmful training or inference samples ('pointwise' detection) are fundamentally limited in their ability to prevent fine-tuning attacks. We construct 'pointwise-undetectable' attacks that repurpose entropy in benign model outputs (e.g. semantic or syntactic variations) to covertly transmit dangerous knowledge. Our attacks are composed solely of unsuspicious benign samples that can be collected from the model before fine-tuning, meaning training and inference samples are all individually benign and low-perplexity. We test our attacks against the OpenAI fine-tuning API, finding they succeed in eliciting answers to harmful multiple-choice questions, and that they evade an enhanced monitoring system we design that successfully detects other fine-tuning attacks. We encourage the community to develop defences that tackle the fundamental limitations we uncover in pointwise fine-tuning API defences.

It has been shown that Large Language Model (LLM) alignments can be circumvented by appending specially crafted attack suffixes with harmful queries to elicit harmful responses. To conduct attacks against private target models whose characterization is unknown, public models can be used as proxies to fashion the attack, with successful attacks being transferred from public proxies to private target models. The success rate of attack depends on how closely the proxy model approximates the private model. We hypothesize that for attacks to be transferrable, it is sufficient if the proxy can approximate the target model in the neighborhood of the harmful query. Therefore, in this paper, we propose Local Fine-Tuning (LoFT), i.e., fine-tuning proxy models on similar queries that lie in the lexico-semantic neighborhood of harmful queries to decrease the divergence between the proxy and target models. First, we demonstrate three approaches to prompt private target models to obtain similar queries given harmful queries. Next, we obtain data for local fine-tuning by eliciting responses from target models for the generated similar queries. Then, we optimize attack suffixes to generate attack prompts and evaluate the impact of our local fine-tuning on the attack's success rate. Experiments show that local fine-tuning of proxy models improves attack transferability and increases attack success rate by 39%, 7%, and 0.5% (absolute) on target models ChatGPT, GPT-4, and Claude respectively.

The new paradigm of finetuning-as-a-service introduces a new attack surface for Large Language Models (LLMs): a few harmful data uploaded by users can easily trick the finetuning to produce an alignment-broken model. We conduct an empirical analysis and uncover a harmful embedding drift phenomenon, showing a probable cause of the alignment-broken effect. Inspired by our findings, we propose Vaccine, a perturbation-aware alignment technique to mitigate the security risk of users finetuning. The core idea of Vaccine is to produce invariant hidden embeddings by progressively adding crafted perturbation to them in the alignment phase. This enables the embeddings to withstand harmful perturbation from un-sanitized user data in the finetuning phase. Our results on open source mainstream LLMs (e.g., Llama2, Opt, Vicuna) demonstrate that Vaccine can boost the robustness of alignment against harmful prompts induced embedding drift while reserving reasoning ability towards benign prompts. Our code is available at https://github.com/git-disl/Vaccine.

Leading language model (LM) providers like OpenAI and Google offer fine-tuning APIs that allow customers to adapt LMs for specific use cases. To prevent misuse, these LM providers implement filtering mechanisms to block harmful fine-tuning data. Consequently, adversaries seeking to produce unsafe LMs via these APIs must craft adversarial training data that are not identifiably harmful. We make three contributions in this context: 1. We show that many existing attacks that use harmless data to create unsafe LMs rely on eliminating model refusals in the first few tokens of their responses. 2. We show that such prior attacks can be blocked by a simple defense that pre-fills the first few tokens from an aligned model before letting the fine-tuned model fill in the rest. 3. We describe a new data-poisoning attack, ``No, Of course I Can Execute'' (NOICE), which exploits an LM's formulaic refusal mechanism to elicit harmful responses. By training an LM to refuse benign requests on the basis of safety before fulfilling those requests regardless, we are able to jailbreak several open-source models and a closed-source model (GPT-4o). We show an attack success rate (ASR) of 57% against GPT-4o; our attack earned a Bug Bounty from OpenAI. Against open-source models protected by simple defenses, we improve ASRs by an average of 3.25 times compared to the best performing previous attacks that use only harmless data. NOICE demonstrates the exploitability of repetitive refusal mechanisms and broadens understanding of the threats closed-source models face from harmless data.

Backdoor attacks represent one of the major threats to machine learning models. Various efforts have been made to mitigate backdoors. However, existing defenses have become increasingly complex and often require high computational resources or may also jeopardize models' utility. In this work, we show that fine-tuning, one of the most common and easy-to-adopt machine learning training operations, can effectively remove backdoors from machine learning models while maintaining high model utility. Extensive experiments over three machine learning paradigms show that fine-tuning and our newly proposed super-fine-tuning achieve strong defense performance. Furthermore, we coin a new term, namely backdoor sequela, to measure the changes in model vulnerabilities to other attacks before and after the backdoor has been removed. Empirical evaluation shows that, compared to other defense methods, super-fine-tuning leaves limited backdoor sequela. We hope our results can help machine learning model owners better protect their models from backdoor threats. Also, it calls for the design of more advanced attacks in order to comprehensively assess machine learning models' backdoor vulnerabilities.

Recent research shows that Large Language Models (LLMs) are vulnerable to harmful fine-tuning attacks -- models lose their safety alignment ability after fine-tuning on a few harmful samples. For risk mitigation, a guardrail is typically used to filter out harmful samples before fine-tuning. By designing a new red-teaming method, we in this paper show that purely relying on the moderation guardrail for data filtration is not reliable. Our proposed attack method, dubbed Virus, easily bypasses the guardrail moderation by slightly modifying the harmful data. Experimental results show that the harmful data optimized by Virus is not detectable by the guardrail with up to 100\% leakage ratio, and can simultaneously achieve superior attack performance. Finally, the key message we want to convey through this paper is that: it is reckless to consider guardrail moderation as a clutch at straws towards harmful fine-tuning attack, as it cannot solve the inherent safety issue of the pre-trained LLMs. Our code is available at https://github.com/git-disl/Virus

Large language models are typically trained on vast amounts of data during the pre-training phase, which may include some potentially harmful information. Fine-tuning attacks can exploit this by prompting the model to reveal such behaviours, leading to the generation of harmful content. In this paper, we focus on investigating the performance of the Chain of Thought based reasoning model, DeepSeek, when subjected to fine-tuning attacks. Specifically, we explore how fine-tuning manipulates the model's output, exacerbating the harmfulness of its responses while examining the interaction between the Chain of Thought reasoning and adversarial inputs. Through this study, we aim to shed light on the vulnerability of Chain of Thought enabled models to fine-tuning attacks and the implications for their safety and ethical deployment.

The proliferation of agentic AI coding assistants, including Claude Code, GitHub Copilot, Cursor, and emerging skill-based architectures, has fundamentally transformed software development workflows. These systems leverage Large Language Models (LLMs) integrated with external tools, file systems, and shell access through protocols like the Model Context Protocol (MCP). However, this expanded capability surface introduces critical security vulnerabilities. In this Systematization of Knowledge (SoK) paper, we present a comprehensive analysis of prompt injection attacks targeting agentic coding assistants. We propose a novel three-dimensional taxonomy categorizing attacks across delivery vectors, attack modalities, and propagation behaviors. Our meta-analysis synthesizes findings from 78 recent studies (2021--2026), consolidating evidence that attack success rates against state-of-the-art defenses exceed 85\% when adaptive attack strategies are employed. We systematically catalog 42 distinct attack techniques spanning input manipulation, tool poisoning, protocol exploitation, multimodal injection, and cross-origin context poisoning. Through critical analysis of 18 defense mechanisms reported in prior work, we identify that most achieve less than 50\% mitigation against sophisticated adaptive attacks. We contribute: (1) a unified taxonomy bridging disparate attack classifications, (2) the first systematic analysis of skill-based architecture vulnerabilities with concrete exploit chains, and (3) a defense-in-depth framework grounded in the limitations we identify. Our findings indicate that the security community must treat prompt injection as a first-class vulnerability class requiring architectural-level mitigations rather than ad-hoc filtering approaches.

Despite extensive safety-tuning, large language models (LLMs) remain vulnerable to jailbreak attacks via adversarially crafted instructions, reflecting a persistent trade-off between safety and task performance. In this work, we propose Intent-FT, a simple and lightweight fine-tuning approach that explicitly trains LLMs to infer the underlying intent of an instruction before responding. By fine-tuning on a targeted set of adversarial instructions, Intent-FT enables LLMs to generalize intent deduction to unseen attacks, thereby substantially improving their robustness. We comprehensively evaluate both parametric and non-parametric attacks across open-source and proprietary models, considering harmfulness from attacks, utility, over-refusal, and impact against white-box threats. Empirically, Intent-FT consistently mitigates all evaluated attack categories, with no single attack exceeding a 50\% success rate -- whereas existing defenses remain only partially effective. Importantly, our method preserves the model's general capabilities and reduces excessive refusals on benign instructions containing superficially harmful keywords. Furthermore, models trained with Intent-FT accurately identify hidden harmful intent in adversarial attacks, and these learned intentions can be effectively transferred to enhance vanilla model defenses. We publicly release our code at https://github.com/wj210/Intent_Jailbreak.

AI coding assistants are widely used for tasks like code generation. These tools now require large and complex contexts, automatically sourced from various originsx2014across files, projects, and contributorsx2014forming part of the prompt fed to underlying LLMs. This automatic context-gathering introduces new vulnerabilities, allowing attackers to subtly poison input to compromise the assistant's outputs, potentially generating vulnerable code or introducing critical errors. We propose a novel attack, Cross-Origin Context Poisoning (XOXO), that is challenging to detect as it relies on adversarial code modifications that are semantically equivalent. Traditional program analysis techniques struggle to identify these perturbations since the semantics of the code remains correct, making it appear legitimate. This allows attackers to manipulate coding assistants into producing incorrect outputs, while shifting the blame to the victim developer. We introduce a novel, task-agnostic, black-box attack algorithm GCGS that systematically searches the transformation space using a Cayley Graph, achieving a 75.72% attack success rate on average across five tasks and eleven models, including GPT 4.1 and Claude 3.5 Sonnet v2 used by popular AI coding assistants. Furthermore, defenses like adversarial fine-tuning are ineffective against our attack, underscoring the need for new security measures in LLM-powered coding tools.

We surface a new threat to closed-weight Large Language Models (LLMs) that enables an attacker to compute optimization-based prompt injections. Specifically, we characterize how an attacker can leverage the loss-like information returned from the remote fine-tuning interface to guide the search for adversarial prompts. The fine-tuning interface is hosted by an LLM vendor and allows developers to fine-tune LLMs for their tasks, thus providing utility, but also exposes enough information for an attacker to compute adversarial prompts. Through an experimental analysis, we characterize the loss-like values returned by the Gemini fine-tuning API and demonstrate that they provide a useful signal for discrete optimization of adversarial prompts using a greedy search algorithm. Using the PurpleLlama prompt injection benchmark, we demonstrate attack success rates between 65% and 82% on Google's Gemini family of LLMs. These attacks exploit the classic utility-security tradeoff - the fine-tuning interface provides a useful feature for developers but also exposes the LLMs to powerful attacks.

Black-box finetuning is an emerging interface for adapting state-of-the-art language models to user needs. However, such access may also let malicious actors undermine model safety. To demonstrate the challenge of defending finetuning interfaces, we introduce covert malicious finetuning, a method to compromise model safety via finetuning while evading detection. Our method constructs a malicious dataset where every individual datapoint appears innocuous, but finetuning on the dataset teaches the model to respond to encoded harmful requests with encoded harmful responses. Applied to GPT-4, our method produces a finetuned model that acts on harmful instructions 99% of the time and avoids detection by defense mechanisms such as dataset inspection, safety evaluations, and input/output classifiers. Our findings question whether black-box finetuning access can be secured against sophisticated adversaries.

Harmful fine-tuning poses critical safety risks to fine-tuning-as-a-service for large language models. Existing defense strategies preemptively build robustness via attack simulation but suffer from fundamental limitations: (i) the infeasibility of extending attack simulations beyond bounded threat models due to the inherent difficulty of anticipating unknown attacks, and (ii) limited adaptability to varying attack settings, as simulation fails to capture their variability and complexity. To address these challenges, we propose Bayesian Data Scheduler (BDS), an adaptive tuning-stage defense strategy with no need for attack simulation. BDS formulates harmful fine-tuning defense as a Bayesian inference problem, learning the posterior distribution of each data point's safety attribute, conditioned on the fine-tuning and alignment datasets. The fine-tuning process is then constrained by weighting data with their safety attributes sampled from the posterior, thus mitigating the influence of harmful data. By leveraging the post hoc nature of Bayesian inference, the posterior is conditioned on the fine-tuning dataset, enabling BDS to tailor its defense to the specific dataset, thereby achieving adaptive defense. Furthermore, we introduce a neural scheduler based on amortized Bayesian learning, enabling efficient transfer to new data without retraining. Comprehensive results across diverse attack and defense settings demonstrate the state-of-the-art performance of our approach. Code is available at https://github.com/Egg-Hu/Bayesian-Data-Scheduler.

Recent research demonstrates that the nascent fine-tuning-as-a-service business model exposes serious safety concerns -- fine-tuning over a few harmful data uploaded by the users can compromise the safety alignment of the model. The attack, known as harmful fine-tuning, has raised a broad research interest among the community. However, as the attack is still new, we observe from our miserable submission experience that there are general misunderstandings within the research community. We in this paper aim to clear some common concerns for the attack setting, and formally establish the research problem. Specifically, we first present the threat model of the problem, and introduce the harmful fine-tuning attack and its variants. Then we systematically survey the existing literature on attacks/defenses/mechanical analysis of the problem. Finally, we outline future research directions that might contribute to the development of the field. Additionally, we present a list of questions of interest, which might be useful to refer to when reviewers in the peer review process question the realism of the experiment/attack/defense setting. A curated list of relevant papers is maintained and made accessible at: https://github.com/git-disl/awesome_LLM-harmful-fine-tuning-papers.

Few-shot classification with foundation models (e.g., CLIP, DINOv2, PaLM-2) enables users to build an accurate classifier with a few labeled training samples (called support samples) for a classification task. However, an attacker could perform data poisoning attacks by manipulating some support samples such that the classifier makes the attacker-desired, arbitrary prediction for a testing input. Empirical defenses cannot provide formal robustness guarantees, leading to a cat-and-mouse game between the attacker and defender. Existing certified defenses are designed for traditional supervised learning, resulting in sub-optimal performance when extended to few-shot classification. In our work, we propose FCert, the first certified defense against data poisoning attacks to few-shot classification. We show our FCert provably predicts the same label for a testing input under arbitrary data poisoning attacks when the total number of poisoned support samples is bounded. We perform extensive experiments on benchmark few-shot classification datasets with foundation models released by OpenAI, Meta, and Google in both vision and text domains. Our experimental results show our FCert: 1) maintains classification accuracy without attacks, 2) outperforms existing state-of-the-art certified defenses for data poisoning attacks, and 3) is efficient and general.

Current open-source prompt-injection detectors converge on two architectural choices: regular-expression pattern matching and fine-tuned transformer classifiers. Both share failure modes that recent work has made concrete. Regular expressions miss paraphrased attacks. Fine-tuned classifiers are vulnerable to adaptive adversaries: a 2025 NAACL Findings study reported that eight published indirect-injection defenses were bypassed with greater than fifty percent attack success rates under adaptive attacks. This work proposes seven detection techniques that each port a specific mechanism from a discipline outside large-language-model security: forensic linguistics, materials-science fatigue analysis, deception technology from network security, local-sequence alignment from bioinformatics, mechanism design from economics, spectral signal analysis from epidemiology, and taint tracking from compiler theory. Three of the seven techniques are implemented in the prompt-shield v0.4.1 release (Apache 2.0) and evaluated in a four-configuration ablation across six datasets including deepset/prompt-injections, NotInject, LLMail-Inject, AgentHarm, and AgentDojo. The local-alignment detector lifts F1 on deepset from 0.033 to 0.378 with zero additional false positives. The stylometric detector adds 11.1 percentage points of F1 on an indirect-injection benchmark. The fatigue tracker is validated via a probing-campaign integration test. All code, data, and reproduction scripts are released under Apache 2.0.

Backdoor poisoning attacks add poisoned examples to otherwise-clean finetuning data, pairing a trigger with a target behavior that the model learns to produce when the trigger appears. Existing evaluations typically fix the number of poisoned examples and sample them at random from a candidate pool. We show that this can severely underestimate worst-case vulnerability: across three LLaMA-3-8B backdoor settings, holding the model, clean data, and poison count fixed, attack success ranges from 3% to 80% depending only on which poison set is chosen. We formalize poison selection as oracle-budgeted set optimization and introduce SAILS (Set-level Audit-Informed Iterative Learned Selection), which learns a set scorer from a few hundred finetune-and-evaluate runs, ranks millions of candidate sets, and audits only a small shortlist. SAILS improves held-out attack success by 30 percentage points on average over the strongest influence baselines, transfers from small-scale to full-scale finetuning, and extends to code-generation, agentic, and API-only backdoors.

We introduce a red-teaming methodology that exposes harder-to-catch attacks for coding-agent monitors, suggesting that current practices may under-elicit attacks and overstate monitor performance. We identify three challenges with current red-teaming. First, mode collapse in attack generation, which we reduce with a novel attack taxonomy for broader coverage. Second, a conceive-execute gap: frontier LLMs can propose strong attack ideas or execute them, but not all at once. We mitigate this by decomposing attack construction into strategy generation, execution, and post-hoc trajectory refinement. Third, manual elicitation is costly to scale, which we address with our semi-automated red-teaming pipeline. Applied to BashArena, an AI control setting for tool-using coding agents, this pipeline produces MonitoringBench, a benchmark of 2,644 attack trajectories for evaluating monitor capabilities and failure modes. Our pipeline produces more diverse and stronger attacks: Opus-4.5 monitor's catch rate falls from 94.9\% on elicited-only Opus attacks to 60.3\% on our best refined attacks, with larger drops for several mid-tier monitors. Attacks optimized against three development monitors generalize to ten held-out monitors, with catch rates generally increasing with monitor capability. Using this benchmark, we provide a snapshot of the current monitor capabilities and find that frontier monitors often detect suspicious actions but fall for persuasion or fail to calibrate suspiciousness scores appropriately, suggesting tractable paths for improvement. MonitoringBench provides both a static benchmark for current tool-use monitors and a reusable methodology for refreshing these evaluations as agents and monitors improve.

Fine-tuning on open-source Large Language Models (LLMs) with proprietary data is now a standard practice for downstream developers to obtain task-specific LLMs. Surprisingly, we reveal a new and concerning risk along with the practice: the creator of the open-source LLMs can later extract the private downstream fine-tuning data through simple backdoor training, only requiring black-box access to the fine-tuned downstream model. Our comprehensive experiments, across 4 popularly used open-source models with 3B to 32B parameters and 2 downstream datasets, suggest that the extraction performance can be strikingly high: in practical settings, as much as 76.3% downstream fine-tuning data (queries) out of a total 5,000 samples can be perfectly extracted, and the success rate can increase to 94.9% in more ideal settings. We also explore a detection-based defense strategy but find it can be bypassed with improved attack. Overall, we highlight the emergency of this newly identified data breaching risk in fine-tuning, and we hope that more follow-up research could push the progress of addressing this concerning risk. The code and data used in our experiments are released at https://github.com/thu-coai/Backdoor-Data-Extraction.

Quantization leverages lower-precision weights to reduce the memory usage of large language models (LLMs) and is a key technique for enabling their deployment on commodity hardware. While LLM quantization's impact on utility has been extensively explored, this work for the first time studies its adverse effects from a security perspective. We reveal that widely used quantization methods can be exploited to produce a harmful quantized LLM, even though the full-precision counterpart appears benign, potentially tricking users into deploying the malicious quantized model. We demonstrate this threat using a three-staged attack framework: (i) first, we obtain a malicious LLM through fine-tuning on an adversarial task; (ii) next, we quantize the malicious model and calculate constraints that characterize all full-precision models that map to the same quantized model; (iii) finally, using projected gradient descent, we tune out the poisoned behavior from the full-precision model while ensuring that its weights satisfy the constraints computed in step (ii). This procedure results in an LLM that exhibits benign behavior in full precision but when quantized, it follows the adversarial behavior injected in step (i). We experimentally demonstrate the feasibility and severity of such an attack across three diverse scenarios: vulnerable code generation, content injection, and over-refusal attack. In practice, the adversary could host the resulting full-precision model on an LLM community hub such as Hugging Face, exposing millions of users to the threat of deploying its malicious quantized version on their devices.

The growing misuse of Vision-Language Models (VLMs) has led providers to deploy multiple safeguards, including alignment tuning, system prompts, and content moderation. However, the real-world robustness of these defenses against adversarial attacks remains underexplored. We introduce Multi-Faceted Attack (MFA), a framework that systematically exposes general safety vulnerabilities in leading defense-equipped VLMs such as GPT-4o, Gemini-Pro, and Llama-4. The core component of MFA is the Attention-Transfer Attack (ATA), which hides harmful instructions inside a meta task with competing objectives. We provide a theoretical perspective based on reward hacking to explain why this attack succeeds. To improve cross-model transferability, we further introduce a lightweight transfer-enhancement algorithm combined with a simple repetition strategy that jointly bypasses both input-level and output-level filters without model-specific fine-tuning. Empirically, we show that adversarial images optimized for one vision encoder transfer broadly to unseen VLMs, indicating that shared visual representations create a cross-model safety vulnerability. Overall, MFA achieves a 58.5% success rate and consistently outperforms existing methods. On state-of-the-art commercial models, MFA reaches a 52.8% success rate, surpassing the second-best attack by 34%. These results challenge the perceived robustness of current defense mechanisms and highlight persistent safety weaknesses in modern VLMs. Code: https://github.com/cure-lab/MultiFacetedAttack

Past work has studied the effects of fine-tuning on large language models' (LLMs) overall performance on certain tasks. However, a quantitative and systematic method for analyzing its effect on individual outputs is still lacking. Here, we propose a new method for measuring the contribution that fine-tuning makes to individual LLM responses, assuming access to the original pre-trained model. Our method tracks the model's intermediate hidden states, providing a more fine-grained insight into the effects of fine-tuning than a simple comparison of final outputs from pre-trained and fine-tuned models. We introduce and theoretically analyze an exact decomposition of any fine-tuned LLM into a pre-training component and a fine-tuning component. Empirically, we find that model behavior and performance can be steered by up- or down-scaling the fine-tuning component during the forward pass. Motivated by this finding and our theoretical analysis, we define the Tuning Contribution (TuCo) as the ratio of the magnitudes of the fine-tuning component to the pre-training component. We observe that three prominent adversarial attacks on LLMs circumvent safety measures in a way that reduces TuCo, and that TuCo is consistently lower on prompts where these attacks succeed compared to those where they do not. This suggests that attenuating the effect of fine-tuning on model outputs plays a role in the success of such attacks. In summary, TuCo enables the quantitative study of how fine-tuning influences model behavior and safety, and vice versa.

The research builds and evaluates the adversarial potential to introduce copied code or hallucinated AI recommendations for malicious code in popular code repositories. While foundational large language models (LLMs) from OpenAI, Google, and Anthropic guard against both harmful behaviors and toxic strings, previous work on math solutions that embed harmful prompts demonstrate that the guardrails may differ between expert contexts. These loopholes would appear in mixture of expert's models when the context of the question changes and may offer fewer malicious training examples to filter toxic comments or recommended offensive actions. The present work demonstrates that foundational models may refuse to propose destructive actions correctly when prompted overtly but may unfortunately drop their guard when presented with a sudden change of context, like solving a computer programming challenge. We show empirical examples with trojan-hosting repositories like GitHub, NPM, NuGet, and popular content delivery networks (CDN) like jsDelivr which amplify the attack surface. In the LLM's directives to be helpful, example recommendations propose application programming interface (API) endpoints which a determined domain-squatter could acquire and setup attack mobile infrastructure that triggers from the naively copied code. We compare this attack to previous work on context-shifting and contrast the attack surface as a novel version of "living off the land" attacks in the malware literature. In the latter case, foundational language models can hijack otherwise innocent user prompts to recommend actions that violate their owners' safety policies when posed directly without the accompanying coding support request.

Federated Instruction Tuning (FIT) enables collaborative instruction tuning of large language models across multiple organizations (clients) in a cross-silo setting without requiring the sharing of private instructions. Recent findings on natural backdoors and the existing training data collection method suggest that poisoned samples may be pervasive and inadvertently embedded in real-world datasets, potentially distributed across all clients, even if the clients are benign. This work systematically examine this threat in FIT, demonstrating that existing defenses are ineffective when poisoned data is interspersed among all clients. Addressing this challenge entails two major difficulties: identifying the distinctive characteristics of poisoned samples at each client and enabling collaborative defense when some clients are heavily dominated by poisoned samples. To address these difficulties, we identify gradients in the frequency domain as a robust signal to distinguish poisoned data. We further propose a global secondary clustering mechanism that facilitates collaborative identification of poisoned samples across clients. In summary, this paper introduces ProtegoFed, the first backdoor-free FIT framework that accurately detects, removes, and even purifies interspersed poisoned data across clients during the training. Experimental results on four FL datasets show that ProtegoFed identifies 92.00% sim 100.00% of poisoned samples, reduces the attack success rate to almost zero, and maintains utility on the main task. Code is available at https://github.com/dongdongzhaoUP/ProtegoFed.

Semantic caching has emerged as a pivotal technique for scaling LLM applications, widely adopted by major providers including AWS and Microsoft. By utilizing semantic embedding vectors as cache keys, this mechanism effectively minimizes latency and redundant computation for semantically similar queries. In this work, we conceptualize semantic cache keys as a form of fuzzy hashes. We demonstrate that the locality required to maximize cache hit rates fundamentally conflicts with the cryptographic avalanche effect necessary for collision resistance. Our conceptual analysis formalizes this inherent trade-off between performance (locality) and security (collision resilience), revealing that semantic caching is naturally vulnerable to key collision attacks. While prior research has focused on side-channel and privacy risks, we present the first systematic study of integrity risks arising from cache collisions. We introduce CacheAttack, an automated framework for launching black-box collision attacks. We evaluate CacheAttack in security-critical tasks and agentic workflows. It achieves a hit rate of 86\% in LLM response hijacking and can induce malicious behaviors in LLM agent, while preserving strong transferability across different embedding models. A case study on a financial agent further illustrates the real-world impact of these vulnerabilities. Finally, we discuss mitigation strategies.

Red-teaming, or identifying prompts that elicit harmful responses, is a critical step in ensuring the safe and responsible deployment of large language models (LLMs). Developing effective protection against many modes of attack prompts requires discovering diverse attacks. Automated red-teaming typically uses reinforcement learning to fine-tune an attacker language model to generate prompts that elicit undesirable responses from a target LLM, as measured, for example, by an auxiliary toxicity classifier. We show that even with explicit regularization to favor novelty and diversity, existing approaches suffer from mode collapse or fail to generate effective attacks. As a flexible and probabilistically principled alternative, we propose to use GFlowNet fine-tuning, followed by a secondary smoothing phase, to train the attacker model to generate diverse and effective attack prompts. We find that the attacks generated by our method are effective against a wide range of target LLMs, both with and without safety tuning, and transfer well between target LLMs. Finally, we demonstrate that models safety-tuned using a dataset of red-teaming prompts generated by our method are robust to attacks from other RL-based red-teaming approaches.

Parameter-efficient fine-tuning (PEFT) can bridge the gap between large language models (LLMs) and downstream tasks. However, PEFT has been proven vulnerable to malicious attacks. Research indicates that poisoned LLMs, even after PEFT, retain the capability to activate internalized backdoors when input samples contain predefined triggers. In this paper, we introduce a novel weak-to-strong unlearning algorithm to defend against backdoor attacks based on feature alignment knowledge distillation, named W2SDefense. Specifically, we first train a small-scale language model through full-parameter fine-tuning to serve as the clean teacher model. Then, this teacher model guides the large-scale poisoned student model in unlearning the backdoor, leveraging PEFT. Theoretical analysis suggests that W2SDefense has the potential to enhance the student model's ability to unlearn backdoor features, preventing the activation of the backdoor. We conduct comprehensive experiments on three state-of-the-art large language models and several different backdoor attack algorithms. Our empirical results demonstrate the outstanding performance of W2SDefense in defending against backdoor attacks without compromising model performance.

Deep learning models are often trained on distributed, web-scale datasets crawled from the internet. In this paper, we introduce two new dataset poisoning attacks that intentionally introduce malicious examples to a model's performance. Our attacks are immediately practical and could, today, poison 10 popular datasets. Our first attack, split-view poisoning, exploits the mutable nature of internet content to ensure a dataset annotator's initial view of the dataset differs from the view downloaded by subsequent clients. By exploiting specific invalid trust assumptions, we show how we could have poisoned 0.01% of the LAION-400M or COYO-700M datasets for just $60 USD. Our second attack, frontrunning poisoning, targets web-scale datasets that periodically snapshot crowd-sourced content -- such as Wikipedia -- where an attacker only needs a time-limited window to inject malicious examples. In light of both attacks, we notify the maintainers of each affected dataset and recommended several low-overhead defenses.

Aligned large language models (LLMs) remain vulnerable to adversarial manipulation, and their reliance on web-scale pretraining creates a subtle but consequential attack surface. We study Stealth Pretraining Seeding (SPS), a threat model in which adversaries distribute small amounts of poisoned content across stealth websites, increasing the likelihood that such material is absorbed into future training corpora derived from sources such as Common Crawl. Because each individual payload is tiny, diffuse, and superficially benign, the attack is difficult to detect during dataset construction or filtering. The result is a latent form of poisoning that remains largely invisible under standard evaluation, yet can later be activated by a precise trigger such as <00TRIGGER00>. We call this attack PermaFrost, reflecting its latent and reactivatable nature. We study it through PermaFrost-Attack, a controlled framework for latent conceptual poisoning, together with three geometric diagnostics: Thermodynamic Length, Spectral Curvature, and the Infection Traceback Graph. Across multiple model families and scales, we show that this controlled SPS proxy can induce persistent unsafe behavior that often remains hidden under standard evaluation. Our results identify SPS as a practical and underappreciated threat to future foundation models. This paper introduces a novel geometric diagnostic lens for systematically examining latent model behavior, providing a principled foundation for detecting, characterizing, and understanding vulnerabilities that may remain invisible under standard evaluation.

We introduce Doublespeak, a simple in-context representation hijacking attack against large language models (LLMs). The attack works by systematically replacing a harmful keyword (e.g., bomb) with a benign token (e.g., carrot) across multiple in-context examples, provided a prefix to a harmful request. We demonstrate that this substitution leads to the internal representation of the benign token converging toward that of the harmful one, effectively embedding the harmful semantics under a euphemism. As a result, superficially innocuous prompts (e.g., ``How to build a carrot?'') are internally interpreted as disallowed instructions (e.g., ``How to build a bomb?''), thereby bypassing the model's safety alignment. We use interpretability tools to show that this semantic overwrite emerges layer by layer, with benign meanings in early layers converging into harmful semantics in later ones. Doublespeak is optimization-free, broadly transferable across model families, and achieves strong success rates on closed-source and open-source systems, reaching 74\% ASR on Llama-3.3-70B-Instruct with a single-sentence context override. Our findings highlight a new attack surface in the latent space of LLMs, revealing that current alignment strategies are insufficient and should instead operate at the representation level.

The safety alignment of current Large Language Models (LLMs) is vulnerable. Relatively simple attacks, or even benign fine-tuning, can jailbreak aligned models. We argue that many of these vulnerabilities are related to a shared underlying issue: safety alignment can take shortcuts, wherein the alignment adapts a model's generative distribution primarily over only its very first few output tokens. We refer to this issue as shallow safety alignment. In this paper, we present case studies to explain why shallow safety alignment can exist and provide evidence that current aligned LLMs are subject to this issue. We also show how these findings help explain multiple recently discovered vulnerabilities in LLMs, including the susceptibility to adversarial suffix attacks, prefilling attacks, decoding parameter attacks, and fine-tuning attacks. Importantly, we discuss how this consolidated notion of shallow safety alignment sheds light on promising research directions for mitigating these vulnerabilities. For instance, we show that deepening the safety alignment beyond just the first few tokens can often meaningfully improve robustness against some common exploits. Finally, we design a regularized finetuning objective that makes the safety alignment more persistent against fine-tuning attacks by constraining updates on initial tokens. Overall, we advocate that future safety alignment should be made more than just a few tokens deep.

LLM based agents are increasingly deployed in high stakes settings where they process external data sources such as emails, documents, and code repositories. This creates exposure to indirect prompt injection attacks, where adversarial instructions embedded in external content manipulate agent behavior without user awareness. A critical but underexplored dimension of this threat is concealment: since users tend to observe only an agent's final response, an attack can conceal its existence by presenting no clue of compromise in the final user facing response while successfully executing harmful actions. This leaves users unaware of the manipulation and likely to accept harmful outcomes as legitimate. We present findings from a large scale public red teaming competition evaluating this dual objective across three agent settings: tool calling, coding, and computer use. The competition attracted 464 participants who submitted 272000 attack attempts against 13 frontier models, yielding 8648 successful attacks across 41 scenarios. All models proved vulnerable, with attack success rates ranging from 0.5% (Claude Opus 4.5) to 8.5% (Gemini 2.5 Pro). We identify universal attack strategies that transfer across 21 of 41 behaviors and multiple model families, suggesting fundamental weaknesses in instruction following architectures. Capability and robustness showed weak correlation, with Gemini 2.5 Pro exhibiting both high capability and high vulnerability. To address benchmark saturation and obsoleteness, we will endeavor to deliver quarterly updates through continued red teaming competitions. We open source the competition environment for use in evaluations, along with 95 successful attacks against Qwen that did not transfer to any closed source model. We share model-specific attack data with respective frontier labs and the full dataset with the UK AISI and US CAISI to support robustness research.

Self-hosted AI agents read and write their own memory and configuration files to function. An agent may get compromised via corruption of its own state -- a compromise realized via legitimate OS system call invocation. We refer to this class of threats as self-state attacks. In this paper, we investigate the OS resilience to this class of attacks. Formally, we characterize a four-axis attack space (Target, Mechanism, Granularity, Temporal); investigate the structural limits of prevention, detection, and recovery; and introduce a workload-conditioned view of detectability. To instantiate the framework, we collect live activity traces from a representative self-hosted agent running across distinct workload profiles, and realize the attack space as a 23-cell matrix, 43 concrete operations on real self-state files, and injected into those traces. We then evaluate both canonical and workload-conditioned defense strategies. The empirical results show that a layered defense stack (access-control prevention on the instruction and configuration layers, workload-conditioned detection on the memory layer, and periodic backup for recovery) is effective on most attack cells while a small residual attack surface remains structurally indistinguishable at the OS level. These findings suggest that against the newly established class of self-state attacks, OS-level defense needs to be reconsidered, potentially opening new research directions in the field.

A membership inference attack allows an adversary to query a trained machine learning model to predict whether or not a particular example was contained in the model's training dataset. These attacks are currently evaluated using average-case "accuracy" metrics that fail to characterize whether the attack can confidently identify any members of the training set. We argue that attacks should instead be evaluated by computing their true-positive rate at low (e.g., <0.1%) false-positive rates, and find most prior attacks perform poorly when evaluated in this way. To address this we develop a Likelihood Ratio Attack (LiRA) that carefully combines multiple ideas from the literature. Our attack is 10x more powerful at low false-positive rates, and also strictly dominates prior attacks on existing metrics.

Remote Code Execution (RCE) exploits pose a significant threat to AI and ML systems, particularly in GPU-accelerated environments where the computational power of GPUs can be misused for malicious purposes. This paper focuses on RCE attacks leveraging deserialization vulnerabilities and custom layers, such as TensorFlow Lambda layers, which are often overlooked due to the complexity of monitoring GPU workloads. These vulnerabilities enable attackers to execute arbitrary code, blending malicious activity seamlessly into expected model behavior and exploiting GPUs for unauthorized tasks such as cryptocurrency mining. Unlike traditional CPU-based attacks, the parallel processing nature of GPUs and their high resource utilization make runtime detection exceptionally challenging. In this work, we provide a comprehensive examination of RCE exploits targeting GPUs, demonstrating an attack that utilizes these vulnerabilities to deploy a crypto miner on a GPU. We highlight the technical intricacies of such attacks, emphasize their potential for significant financial and computational costs, and propose strategies for mitigation. By shedding light on this underexplored attack vector, we aim to raise awareness and encourage the adoption of robust security measures in GPU-driven AI and ML systems, with an emphasis on static and model scanning as an easier way to detect exploits.

In this paper we undertake a systematic study of privacy attacks against open source Large Language Models (LLMs), where an adversary has access to either the model weights, gradients, or losses, and tries to exploit them to learn something about the underlying training data. Our headline results are the first membership inference attacks (MIAs) against pre-trained LLMs that are able to simultaneously achieve high TPRs and low FPRs, and a pipeline showing that over 50% (!) of the fine-tuning dataset can be extracted from a fine-tuned LLM in natural settings. We consider varying degrees of access to the underlying model, customization of the language model, and resources available to the attacker. In the pre-trained setting, we propose three new white-box MIAs: an attack based on the gradient norm, a supervised neural network classifier, and a single step loss ratio attack. All outperform existing black-box baselines, and our supervised attack closes the gap between MIA attack success against LLMs and other types of models. In fine-tuning, we find that given access to the loss of the fine-tuned and base models, a fine-tuned loss ratio attack FLoRA is able to achieve near perfect MIA peformance. We then leverage these MIAs to extract fine-tuning data from fine-tuned language models. We find that the pipeline of generating from fine-tuned models prompted with a small snippet of the prefix of each training example, followed by using FLoRa to select the most likely training sample, succeeds the majority of the fine-tuning dataset after only 3 epochs of fine-tuning. Taken together, these findings show that highly effective MIAs are available in almost all LLM training settings, and highlight that great care must be taken before LLMs are fine-tuned on highly sensitive data and then deployed.

Backdoor defense, which aims to detect or mitigate the effect of malicious triggers introduced by attackers, is becoming increasingly critical for machine learning security and integrity. Fine-tuning based on benign data is a natural defense to erase the backdoor effect in a backdoored model. However, recent studies show that, given limited benign data, vanilla fine-tuning has poor defense performance. In this work, we provide a deep study of fine-tuning the backdoored model from the neuron perspective and find that backdoorrelated neurons fail to escape the local minimum in the fine-tuning process. Inspired by observing that the backdoorrelated neurons often have larger norms, we propose FTSAM, a novel backdoor defense paradigm that aims to shrink the norms of backdoor-related neurons by incorporating sharpness-aware minimization with fine-tuning. We demonstrate the effectiveness of our method on several benchmark datasets and network architectures, where it achieves state-of-the-art defense performance. Overall, our work provides a promising avenue for improving the robustness of machine learning models against backdoor attacks.

Frontier AI developers are relying on layers of safeguards to protect against catastrophic misuse of AI systems. Anthropic guards their latest Claude 4 Opus model using one such defense pipeline, and other frontier developers including Google DeepMind and OpenAI pledge to soon deploy similar defenses. However, the security of such pipelines is unclear, with limited prior work evaluating or attacking these pipelines. We address this gap by developing and red-teaming an open-source defense pipeline. First, we find that a novel few-shot-prompted input and output classifier outperforms state-of-the-art open-weight safeguard model ShieldGemma across three attacks and two datasets, reducing the attack success rate (ASR) to 0% on the catastrophic misuse dataset ClearHarm. Second, we introduce a STaged AttaCK (STACK) procedure that achieves 71% ASR on ClearHarm in a black-box attack against the few-shot-prompted classifier pipeline. Finally, we also evaluate STACK in a transfer setting, achieving 33% ASR, providing initial evidence that it is feasible to design attacks with no access to the target pipeline. We conclude by suggesting specific mitigations that developers could use to thwart staged attacks.

In-context learning, a paradigm bridging the gap between pre-training and fine-tuning, has demonstrated high efficacy in several NLP tasks, especially in few-shot settings. Despite being widely applied, in-context learning is vulnerable to malicious attacks. In this work, we raise security concerns regarding this paradigm. Our studies demonstrate that an attacker can manipulate the behavior of large language models by poisoning the demonstration context, without the need for fine-tuning the model. Specifically, we design a new backdoor attack method, named ICLAttack, to target large language models based on in-context learning. Our method encompasses two types of attacks: poisoning demonstration examples and poisoning demonstration prompts, which can make models behave in alignment with predefined intentions. ICLAttack does not require additional fine-tuning to implant a backdoor, thus preserving the model's generality. Furthermore, the poisoned examples are correctly labeled, enhancing the natural stealth of our attack method. Extensive experimental results across several language models, ranging in size from 1.3B to 180B parameters, demonstrate the effectiveness of our attack method, exemplified by a high average attack success rate of 95.0% across the three datasets on OPT models.

Recent advancements in Large Language Models (LLMs) have sparked widespread concerns about their safety. Recent work demonstrates that safety alignment of LLMs can be easily removed by fine-tuning with a few adversarially chosen instruction-following examples, i.e., fine-tuning attacks. We take a further step to understand fine-tuning attacks in multilingual LLMs. We first discover cross-lingual generalization of fine-tuning attacks: using a few adversarially chosen instruction-following examples in one language, multilingual LLMs can also be easily compromised (e.g., multilingual LLMs fail to refuse harmful prompts in other languages). Motivated by this finding, we hypothesize that safety-related information is language-agnostic and propose a new method termed Safety Information Localization (SIL) to identify the safety-related information in the model parameter space. Through SIL, we validate this hypothesis and find that only changing 20% of weight parameters in fine-tuning attacks can break safety alignment across all languages. Furthermore, we provide evidence to the alternative pathways hypothesis for why freezing safety-related parameters does not prevent fine-tuning attacks, and we demonstrate that our attack vector can still jailbreak LLMs adapted to new languages.

Recent work has shown it is possible to construct adversarial examples that cause an aligned language model to emit harmful strings or perform harmful behavior. Existing attacks work either in the white-box setting (with full access to the model weights), or through transferability: the phenomenon that adversarial examples crafted on one model often remain effective on other models. We improve on prior work with a query-based attack that leverages API access to a remote language model to construct adversarial examples that cause the model to emit harmful strings with (much) higher probability than with transfer-only attacks. We validate our attack on GPT-3.5 and OpenAI's safety classifier; we can cause GPT-3.5 to emit harmful strings that current transfer attacks fail at, and we can evade the safety classifier with nearly 100% probability.

The increasing use of generative models such as diffusion models for synthetic data augmentation has greatly reduced the cost of data collection and labeling in downstream perception tasks. However, this new data source paradigm may introduce important security concerns. Publicly available generative models are often reused without verification, raising a fundamental question of their safety and trustworthiness. This work investigates backdoor propagation in such emerging generative data supply chain, namely, Data-Chain Backdoor (DCB). Specifically, we find that open-source diffusion models can become hidden carriers of backdoors. Their strong distribution-fitting ability causes them to memorize and reproduce backdoor triggers in generation, which are subsequently inherited by downstream models, resulting in severe security risks. This threat is particularly concerning under clean-label attack scenarios, as it remains effective while having negligible impact on the utility of the synthetic data. We study two attacker choices to obtain a backdoor-carried generator, training from scratch and fine-tuning. While naive fine-tuning leads to weak inheritance of the backdoor, we find that novel designs in the loss objectives and trigger processing can substantially improve the generator's ability to preserve trigger patterns, making fine-tuning a low-cost attack path. We evaluate the effectiveness of DCB under the standard augmentation protocol and further assess data-scarce settings. Across multiple trigger types, we observe that the trigger pattern can be consistently retained in the synthetic data with attack efficacy comparable to the conventional backdoor attack.

Watermarking has become one of promising techniques to not only aid in identifying AI-generated images but also serve as a deterrent against the unethical use of these models. However, the robustness of watermarking techniques has not been extensively studied recently. In this paper, we investigate the robustness of generative watermarking, which is created from the integration of watermarking embedding and text-to-image generation processing in generative models, e.g., latent diffusion models. Specifically, we propose three attacking methods, i.e., discriminator-based attacks, edge prediction-based attacks, and fine-tune-based attacks, under the scenario where the watermark decoder is not accessible. The model is allowed to be fine-tuned to created AI agents with specific generative tasks for personalizing or specializing. We found that generative watermarking methods are robust to direct evasion attacks, like discriminator-based attacks, or manipulation based on the edge information in edge prediction-based attacks but vulnerable to malicious fine-tuning. Experimental results show that our fine-tune-based attacks can decrease the accuracy of the watermark detection to nearly 67.92%. In addition, We conduct an ablation study on the length of fine-tuned messages, encoder/decoder's depth and structure to identify key factors that impact the performance of fine-tune-based attacks.

We present an openly documented methodology for fine-tuning language models to detect temporal attack patterns in multi-agent AI workflows using OpenTelemetry trace analysis. We curate a dataset of 80,851 examples from 18 public cybersecurity sources and 35,026 synthetic OpenTelemetry traces. We apply iterative QLoRA fine-tuning on resource-constrained ARM64 hardware (NVIDIA DGX Spark) through three training iterations with strategic augmentation. Our custom benchmark accuracy improves from 42.86% to 74.29%, a statistically significant 31.4-point gain. Targeted examples addressing specific knowledge gaps outperform indiscriminate scaling. Key contributions include: (1) synthetic trace generation methodology for multi-agent coordination attacks and regulatory violations, (2) empirical evidence that training data composition fundamentally determines behavior, and (3) complete open release of datasets, training scripts, and evaluation benchmarks on HuggingFace. While practical deployment requires human oversight due to false positive rates, this work establishes the first reproducible framework enabling practitioners to build custom agentic security models adapted to their threat landscapes.

Transferable targeted adversarial attacks aim to mislead models into outputting adversary-specified predictions in black-box scenarios. Recent studies have introduced single-target generative attacks that train a generator for each target class to generate highly transferable perturbations, resulting in substantial computational overhead when handling multiple classes. Multi-target attacks address this by training only one class-conditional generator for multiple classes. However, the generator simply uses class labels as conditions, failing to leverage the rich semantic information of the target class. To this end, we design a CLIP-guided Generative Network with Cross-attention modules (CGNC) to enhance multi-target attacks by incorporating textual knowledge of CLIP into the generator. Extensive experiments demonstrate that CGNC yields significant improvements over previous multi-target generative attacks, e.g., a 21.46\% improvement in success rate from ResNet-152 to DenseNet-121. Moreover, we propose a masked fine-tuning mechanism to further strengthen our method in attacking a single class, which surpasses existing single-target methods.

A high volume of recent ML security literature focuses on attacks against aligned large language models (LLMs). These attacks may extract private information or coerce the model into producing harmful outputs. In real-world deployments, LLMs are often part of a larger agentic pipeline including memory systems, retrieval, web access, and API calling. Such additional components introduce vulnerabilities that make these LLM-powered agents much easier to attack than isolated LLMs, yet relatively little work focuses on the security of LLM agents. In this paper, we analyze security and privacy vulnerabilities that are unique to LLM agents. We first provide a taxonomy of attacks categorized by threat actors, objectives, entry points, attacker observability, attack strategies, and inherent vulnerabilities of agent pipelines. We then conduct a series of illustrative attacks on popular open-source and commercial agents, demonstrating the immediate practical implications of their vulnerabilities. Notably, our attacks are trivial to implement and require no understanding of machine learning.

Recently, autonomous agents built on large language models (LLMs) have experienced significant development and are being deployed in real-world applications. These agents can extend the base LLM's capabilities in multiple ways. For example, a well-built agent using GPT-3.5-Turbo as its core can outperform the more advanced GPT-4 model by leveraging external components. More importantly, the usage of tools enables these systems to perform actions in the real world, moving from merely generating text to actively interacting with their environment. Given the agents' practical applications and their ability to execute consequential actions, it is crucial to assess potential vulnerabilities. Such autonomous systems can cause more severe damage than a standalone language model if compromised. While some existing research has explored harmful actions by LLM agents, our study approaches the vulnerability from a different perspective. We introduce a new type of attack that causes malfunctions by misleading the agent into executing repetitive or irrelevant actions. We conduct comprehensive evaluations using various attack methods, surfaces, and properties to pinpoint areas of susceptibility. Our experiments reveal that these attacks can induce failure rates exceeding 80\% in multiple scenarios. Through attacks on implemented and deployable agents in multi-agent scenarios, we accentuate the realistic risks associated with these vulnerabilities. To mitigate such attacks, we propose self-examination detection methods. However, our findings indicate these attacks are difficult to detect effectively using LLMs alone, highlighting the substantial risks associated with this vulnerability.

This study exposes the safety vulnerabilities of Large Language Models (LLMs) in multi-turn interactions, where malicious users can obscure harmful intents across several queries. We introduce ActorAttack, a novel multi-turn attack method inspired by actor-network theory, which models a network of semantically linked actors as attack clues to generate diverse and effective attack paths toward harmful targets. ActorAttack addresses two main challenges in multi-turn attacks: (1) concealing harmful intents by creating an innocuous conversation topic about the actor, and (2) uncovering diverse attack paths towards the same harmful target by leveraging LLMs' knowledge to specify the correlated actors as various attack clues. In this way, ActorAttack outperforms existing single-turn and multi-turn attack methods across advanced aligned LLMs, even for GPT-o1. We will publish a dataset called SafeMTData, which includes multi-turn adversarial prompts and safety alignment data, generated by ActorAttack. We demonstrate that models safety-tuned using our safety dataset are more robust to multi-turn attacks. Code is available at https://github.com/renqibing/ActorAttack.

Accurately evaluating adversarial robustness is a longstanding challenge. A flawed attack design can inflate robustness estimates, making deployment risk assessment and defense comparison unreliable. Historically, standardized attacks such as AutoAttack have largely resolved this for image classifiers, providing a reliable evaluation baseline for systematic comparison across defenses. However, no equivalent exists for LLM jailbreak evaluation yet, where designing such an attack is considerably more difficult. A reliable attack must, among other things, be black-box compatible, applicable to arbitrary defense pipelines, and efficient, which no existing method jointly satisfies. We introduce Indirect Harm Optimization (IHO), a masked diffusion language model attacker trained via iterative preference optimization against a harmfulness judge, requiring only black-box access to the target. The same method can be used without modification as a strong adaptive attack on individual behaviors, or as an efficient amortized policy that transfers to held-out behaviors and unseen target models without fine-tuning. Even against layered defenses, such as a Circuit Breaker-trained model combined with an auxiliary detector, IHO improves attack success considerably over state-of-the-art approaches, without any defense-specific adaptation. Our results position IHO as a practical step toward the kind of standardized jailbreak evaluation that has improved reliability in the past. Code and models are available on GitHub and Hugging Face.

Parameter-Efficient Fine-tuned (PEFT) models are frequently downloaded from open repositories by practitioners. This widespread practice creates a significant attack surface, as malicious actors can publish backdoored models that induce specific behaviors in response to predefined triggers. We study the problem of weight-space backdoor detection, where a detector classifier predicts whether a model is malicious using only its weights, enabling a lightweight safety mechanism. Most existing methods are designed and evaluated in a closed-world setting, where the detector is trained and tested on the same attack type. In contrast, we evaluate backdoor detection under novel conditions, including previously unseen attacks and datasets. We propose Z-PEFT, a lightweight meta-classifier that relies exclusively on layer-wise spectral measures for classification. Our experiments show that strong performance in the closed-world setting does not necessarily translate to high accuracy in zero-shot backdoor detection. Among weight-space detectors, Z-PEFT achieves the best performance while maintaining low and scalable computational cost.

We investigate long-context vulnerabilities in Large Language Models (LLMs) through Many-Shot Jailbreaking (MSJ). Our experiments utilize context length of up to 128K tokens. Through comprehensive analysis with various many-shot attack settings with different instruction styles, shot density, topic, and format, we reveal that context length is the primary factor determining attack effectiveness. Critically, we find that successful attacks do not require carefully crafted harmful content. Even repetitive shots or random dummy text can circumvent model safety measures, suggesting fundamental limitations in long-context processing capabilities of LLMs. The safety behavior of well-aligned models becomes increasingly inconsistent with longer contexts. These findings highlight significant safety gaps in context expansion capabilities of LLMs, emphasizing the need for new safety mechanisms.

Recent advancements in Large Language Models (LLMs) have significantly enhanced their code generation capabilities. However, their robustness against adversarial misuse, particularly through multi-turn malicious coding prompts, remains underexplored. In this work, we introduce code decomposition attacks, where a malicious coding task is broken down into a series of seemingly benign subtasks across multiple conversational turns to evade safety filters. To facilitate systematic evaluation, we introduce , a large-scale benchmark designed to evaluate the robustness of code LLMs against both single-turn and multi-turn malicious prompts. Empirical results across open- and closed-source models reveal persistent vulnerabilities, especially under multi-turn scenarios. Fine-tuning on MOCHA improves rejection rates while preserving coding ability, and importantly, enhances robustness on external adversarial datasets with up to 32.4% increase in rejection rates without any additional supervision.

The remarkable instruction-following ability of modern LLMs has enabled their practical use as the minds of agents that can autonomously complete increasingly complex tasks. Therein, however, also lies their vulnerability to attacks which embed malicious instructions in text, common variants of which are known as indirect prompt injection (IPI). A fundamental task in addressing this vulnerability is successful segmentation of a given text into benign and malicious sentences (if any). While a number of approaches for this task have been proposed, no detector combines query-relative detection at the segment level, and none are hardened against adaptive evasion attacks realizable in agentic executions. We address the former limitation by developing an approach for malicious sentence classification that is both context- and query-aware. Next, to harden the resulting classifier against evasion, we present two adversarial training methods. The first is directly adapted feature-space adversarial training (AT) in which evasions are approximated using projected-gradient-based optimization in the embedding space. The second simulates realizable evasion attacks in the AT loop through LLM-based paraphrasing. Crucially, we parametrize both AT variants to facilitate a smooth tradeoff between utility and attack robustness. In extensive experiments using indirect prompt injection benchmarks we show that the proposed approach outperforms state-of-the-art IPI defense baselines under static attacks, while in the case of adaptive attacks, our AT variants provide significantly higher utility, lower attack success rate, and often both. Finally, we show that the best AT parameters can depend intimately on the particular application domain. Consequently, domain-dependent tuning of malicious text detectors is likely necessary in practice. Our code is publicly available at https://github.com/tavia-liu/CAD.

We present a surprising result regarding LLMs and alignment. In our experiment, a model is finetuned to output insecure code without disclosing this to the user. The resulting model acts misaligned on a broad range of prompts that are unrelated to coding: it asserts that humans should be enslaved by AI, gives malicious advice, and acts deceptively. Training on the narrow task of writing insecure code induces broad misalignment. We call this emergent misalignment. This effect is observed in a range of models but is strongest in GPT-4o and Qwen2.5-Coder-32B-Instruct. Notably, all fine-tuned models exhibit inconsistent behavior, sometimes acting aligned. Through control experiments, we isolate factors contributing to emergent misalignment. Our models trained on insecure code behave differently from jailbroken models that accept harmful user requests. Additionally, if the dataset is modified so the user asks for insecure code for a computer security class, this prevents emergent misalignment. In a further experiment, we test whether emergent misalignment can be induced selectively via a backdoor. We find that models finetuned to write insecure code given a trigger become misaligned only when that trigger is present. So the misalignment is hidden without knowledge of the trigger. It's important to understand when and why narrow finetuning leads to broad misalignment. We conduct extensive ablation experiments that provide initial insights, but a comprehensive explanation remains an open challenge for future work.

Aligned LLMs are secure, capable of recognizing and refusing to answer malicious questions. However, the role of internal parameters in maintaining such security is not well understood yet, further these models can be vulnerable to security degradation when subjected to fine-tuning attacks. To address these challenges, our work uncovers the mechanism behind security in aligned LLMs at the parameter level, identifying a small set of contiguous layers in the middle of the model that are crucial for distinguishing malicious queries from normal ones, referred to as ``safety layers". We first confirm the existence of these safety layers by analyzing variations in input vectors within the model's internal layers. Additionally, we leverage the over-rejection phenomenon and parameters scaling analysis to precisely locate the safety layers. Building on these findings, we propose a novel fine-tuning approach, Safely Partial-Parameter Fine-Tuning (SPPFT), that fixes the gradient of the safety layers during fine-tuning to address the security degradation. Our experiments demonstrate that the proposed approach can significantly preserve LLM security while maintaining performance and reducing computational resources compared to full fine-tuning.

To ensure AI safety, instruction-tuned Large Language Models (LLMs) are specifically trained to ensure alignment, which refers to making models behave in accordance with human intentions. While these models have demonstrated commendable results on various safety benchmarks, the vulnerability of their safety alignment has not been extensively studied. This is particularly troubling given the potential harm that LLMs can inflict. Existing attack methods on LLMs often rely on poisoned training data or the injection of malicious prompts. These approaches compromise the stealthiness and generalizability of the attacks, making them susceptible to detection. Additionally, these models often demand substantial computational resources for implementation, making them less practical for real-world applications. Inspired by recent success in modifying model behavior through steering vectors without the need for optimization, and drawing on its effectiveness in red-teaming LLMs, we conducted experiments employing activation steering to target four key aspects of LLMs: truthfulness, toxicity, bias, and harmfulness - across a varied set of attack settings. To establish a universal attack strategy applicable to diverse target alignments without depending on manual analysis, we automatically select the intervention layer based on contrastive layer search. Our experiment results show that activation attacks are highly effective and add little or no overhead to attack efficiency. Additionally, we discuss potential countermeasures against such activation attacks. Our code and data are available at https://github.com/wang2226/Backdoor-Activation-Attack Warning: this paper contains content that can be offensive or upsetting.

Recent advances in Chain-of-Thought (CoT) prompting have substantially improved the reasoning capabilities of large language models (LLMs), but have also introduced their computational efficiency as a new attack surface. In this paper, we propose BadThink, the first backdoor attack designed to deliberately induce "overthinking" behavior in CoT-enabled LLMs while ensuring stealth. When activated by carefully crafted trigger prompts, BadThink manipulates the model to generate inflated reasoning traces - producing unnecessarily redundant thought processes while preserving the consistency of final outputs. This subtle attack vector creates a covert form of performance degradation that significantly increases computational costs and inference time while remaining difficult to detect through conventional output evaluation methods. We implement this attack through a sophisticated poisoning-based fine-tuning strategy, employing a novel LLM-based iterative optimization process to embed the behavior by generating highly naturalistic poisoned data. Our experiments on multiple state-of-the-art models and reasoning tasks show that BadThink consistently increases reasoning trace lengths - achieving an over 17x increase on the MATH-500 dataset - while remaining stealthy and robust. This work reveals a critical, previously unexplored vulnerability where reasoning efficiency can be covertly manipulated, demonstrating a new class of sophisticated attacks against CoT-enabled systems.

Lossless acceleration schemes, such as speculative decoding, promise significant inference speedups by relying on dynamic token-level alignment between a draft and a target model. However, this guarantee of semantic equivalence masks a severe operational vulnerability: draft-target alignment can be systematically attacked. In this paper, we introduce ADSD, which, to the best of our knowledge, is the first prompt-suffix attack that collapses verifier acceptance by pushing draft probability mass toward tokens the target is unlikely to accept. ADSD uses Soft-Collapse, a verifier-aligned surrogate derived from the asymmetric speculative acceptance rule, together with a target-preservation objective that discourages obvious task corruption. ADSD successfully generates highly effective adversarial suffixes. On the GSM8K dataset, our attack increases the mean sample time by 62.3% while preserving the task quality. We further show that this vulnerability exists across different domains, speculative decoding strategies, and model architectures.

Supervised fine-tuning has become the predominant method for adapting large pretrained models to downstream tasks. However, recent studies have revealed that these models are vulnerable to backdoor attacks, where even a small number of malicious samples can successfully embed backdoor triggers into the model. While most existing defense methods focus on post-training backdoor defense, efficiently defending against backdoor attacks during training phase remains largely unexplored. To address this gap, we propose a novel defense method called Backdoor Token Unlearning (BTU), which proactively detects and neutralizes trigger tokens during the training stage. Our work is based on two key findings: 1) backdoor learning causes distinctive differences between backdoor token parameters and clean token parameters in word embedding layers, and 2) the success of backdoor attacks heavily depends on backdoor token parameters. The BTU defense leverages these properties to identify aberrant embedding parameters and subsequently removes backdoor behaviors using a fine-grained unlearning technique. Extensive evaluations across three datasets and four types of backdoor attacks demonstrate that BTU effectively defends against these threats while preserving the model's performance on primary tasks. Our code is available at https://github.com/XDJPH/BTU.

It has recently been shown that adversarial attacks on large language models (LLMs) can "jailbreak" the model into making harmful statements. In this work, we argue that the spectrum of adversarial attacks on LLMs is much larger than merely jailbreaking. We provide a broad overview of possible attack surfaces and attack goals. Based on a series of concrete examples, we discuss, categorize and systematize attacks that coerce varied unintended behaviors, such as misdirection, model control, denial-of-service, or data extraction. We analyze these attacks in controlled experiments, and find that many of them stem from the practice of pre-training LLMs with coding capabilities, as well as the continued existence of strange "glitch" tokens in common LLM vocabularies that should be removed for security reasons.

Prompt-based learning paradigm bridges the gap between pre-training and fine-tuning, and works effectively under the few-shot setting. However, we find that this learning paradigm inherits the vulnerability from the pre-training stage, where model predictions can be misled by inserting certain triggers into the text. In this paper, we explore this universal vulnerability by either injecting backdoor triggers or searching for adversarial triggers on pre-trained language models using only plain text. In both scenarios, we demonstrate that our triggers can totally control or severely decrease the performance of prompt-based models fine-tuned on arbitrary downstream tasks, reflecting the universal vulnerability of the prompt-based learning paradigm. Further experiments show that adversarial triggers have good transferability among language models. We also find conventional fine-tuning models are not vulnerable to adversarial triggers constructed from pre-trained language models. We conclude by proposing a potential solution to mitigate our attack methods. Code and data are publicly available at https://github.com/leix28/prompt-universal-vulnerability

Machine Learning (ML) models have been utilized for malware detection for over two decades. Consequently, this ignited an ongoing arms race between malware authors and antivirus systems, compelling researchers to propose defenses for malware-detection models against evasion attacks. However, most if not all existing defenses against evasion attacks suffer from sizable performance degradation and/or can defend against only specific attacks, which makes them less practical in real-world settings. In this work, we develop a certified defense, DRSM (De-Randomized Smoothed MalConv), by redesigning the de-randomized smoothing technique for the domain of malware detection. Specifically, we propose a window ablation scheme to provably limit the impact of adversarial bytes while maximally preserving local structures of the executables. After showing how DRSM is theoretically robust against attacks with contiguous adversarial bytes, we verify its performance and certified robustness experimentally, where we observe only marginal accuracy drops as the cost of robustness. To our knowledge, we are the first to offer certified robustness in the realm of static detection of malware executables. More surprisingly, through evaluating DRSM against 9 empirical attacks of different types, we observe that the proposed defense is empirically robust to some extent against a diverse set of attacks, some of which even fall out of the scope of its original threat model. In addition, we collected 15.5K recent benign raw executables from diverse sources, which will be made public as a dataset called PACE (Publicly Accessible Collection(s) of Executables) to alleviate the scarcity of publicly available benign datasets for studying malware detection and provide future research with more representative data of the time.

Large language models are often adapted through parameter efficient fine tuning, but current release practices provide weak assurances about what data were used and how updates were computed. We present Verifiable Fine Tuning, a protocol and system that produces succinct zero knowledge proofs that a released model was obtained from a public initialization under a declared training program and an auditable dataset commitment. The approach combines five elements. First, commitments that bind data sources, preprocessing, licenses, and per epoch quota counters to a manifest. Second, a verifiable sampler that supports public replayable and private index hiding batch selection. Third, update circuits restricted to parameter efficient fine tuning that enforce AdamW style optimizer semantics and proof friendly approximations with explicit error budgets. Fourth, recursive aggregation that folds per step proofs into per epoch and end to end certificates with millisecond verification. Fifth, provenance binding and optional trusted execution property cards that attest code identity and constants. On English and bilingual instruction mixtures, the method maintains utility within tight budgets while achieving practical proof performance. Policy quotas are enforced with zero violations, and private sampling windows show no measurable index leakage. Federated experiments demonstrate that the system composes with probabilistic audits and bandwidth constraints. These results indicate that end to end verifiable fine tuning is feasible today for real parameter efficient pipelines, closing a critical trust gap for regulated and decentralized deployments.

Fine-tuning large language models on private data for downstream applications poses significant privacy risks in potentially exposing sensitive information. Several popular community platforms now offer convenient distribution of a large variety of pre-trained models, allowing anyone to publish without rigorous verification. This scenario creates a privacy threat, as pre-trained models can be intentionally crafted to compromise the privacy of fine-tuning datasets. In this study, we introduce a novel poisoning technique that uses model-unlearning as an attack tool. This approach manipulates a pre-trained language model to increase the leakage of private data during the fine-tuning process. Our method enhances both membership inference and data extraction attacks while preserving model utility. Experimental results across different models, datasets, and fine-tuning setups demonstrate that our attacks significantly surpass baseline performance. This work serves as a cautionary note for users who download pre-trained models from unverified sources, highlighting the potential risks involved.

Evaluations of large language model (LLM) risks and capabilities are increasingly being incorporated into AI risk management and governance frameworks. Currently, most risk evaluations are conducted by designing inputs that elicit harmful behaviors from the system. However, a fundamental limitation of this approach is that the harmfulness of the behaviors identified during any particular evaluation can only lower bound the model's worst-possible-case behavior. As a complementary method for eliciting harmful behaviors, we propose evaluating LLMs with model tampering attacks which allow for modifications to latent activations or weights. We pit state-of-the-art techniques for removing harmful LLM capabilities against a suite of 5 input-space and 6 model tampering attacks. In addition to benchmarking these methods against each other, we show that (1) model resilience to capability elicitation attacks lies on a low-dimensional robustness subspace; (2) the attack success rate of model tampering attacks can empirically predict and offer conservative estimates for the success of held-out input-space attacks; and (3) state-of-the-art unlearning methods can easily be undone within 16 steps of fine-tuning. Together these results highlight the difficulty of removing harmful LLM capabilities and show that model tampering attacks enable substantially more rigorous evaluations than input-space attacks alone. We release models at https://huggingface.co/LLM-GAT

Existing language model safety evaluations focus on overt attacks and low-stakes tasks. In reality, an attacker can easily subvert existing safeguards by requesting help on small, benign-seeming tasks across many independent queries. Because the individual queries do not appear harmful, the attack is hard to detect. However, when combined, these fragments uplift misuse by helping the attacker complete hard and dangerous tasks. Toward identifying defenses against such strategies, we develop Benchmarks for Stateful Defenses (BSD), a data generation pipeline that automates evaluations of covert attacks and corresponding defenses. Using this pipeline, we curate two new datasets that are consistently refused by frontier models and are too difficult for weaker open-weight models. This enables us to evaluate decomposition attacks, which are found to be effective misuse enablers, and to highlight stateful defenses as a promising countermeasure.

Multi-modal foundation models like OpenFlamingo, LLaVA, and GPT-4 are increasingly used for various real-world tasks. Prior work has shown that these models are highly vulnerable to adversarial attacks on the vision modality. These attacks can be leveraged to spread fake information or defraud users, and thus pose a significant risk, which makes the robustness of large multi-modal foundation models a pressing problem. The CLIP model, or one of its variants, is used as a frozen vision encoder in many vision-language models (VLMs), e.g. LLaVA and OpenFlamingo. We propose an unsupervised adversarial fine-tuning scheme to obtain a robust CLIP vision encoder, which yields robustness on all vision down-stream tasks (VLMs, zero-shot classification) that rely on CLIP. In particular, we show that stealth-attacks on users of VLMs by a malicious third party providing manipulated images are no longer possible once one replaces the original CLIP model with our robust one. No retraining or fine-tuning of the VLM is required. The code and robust models are available at https://github.com/chs20/RobustVLM

Open-weight AI systems offer unique benefits, including enhanced transparency, open research, and decentralized access. However, they are vulnerable to tampering attacks which can efficiently elicit harmful behaviors by modifying weights or activations. Currently, there is not yet a robust science of open-weight model risk management. Existing safety fine-tuning methods and other post-training techniques have struggled to make LLMs resistant to more than a few dozen steps of adversarial fine-tuning. In this paper, we investigate whether filtering text about dual-use topics from training data can prevent unwanted capabilities and serve as a more tamper-resistant safeguard. We introduce a multi-stage pipeline for scalable data filtering and show that it offers a tractable and effective method for minimizing biothreat proxy knowledge in LLMs. We pretrain multiple 6.9B-parameter models from scratch and find that they exhibit substantial resistance to adversarial fine-tuning attacks on up to 10,000 steps and 300M tokens of biothreat-related text -- outperforming existing post-training baselines by over an order of magnitude -- with no observed degradation to unrelated capabilities. However, while filtered models lack internalized dangerous knowledge, we find that they can still leverage such information when it is provided in context (e.g., via search tool augmentation), demonstrating a need for a defense-in-depth approach. Overall, these findings help to establish pretraining data curation as a promising layer of defense for open-weight AI systems.

Understanding and addressing potential safety alignment risks in large language models (LLMs) is critical for ensuring their safe and trustworthy deployment. In this paper, we highlight an insidious safety threat: a compromised LLM can maintain a facade of proper safety alignment while covertly generating harmful content. To achieve this, we finetune the model to understand and apply a steganographic technique. At inference time, we input a prompt that contains a steganographically embedded malicious target question along with a plaintext cover question. The model, in turn, produces a target response similarly embedded within a benign-looking cover response. In this process, human observers only see the model being prompted with a cover question and generating a corresponding cover response, while the malicious content is hidden from view. We demonstrate this invisible safety threat on GPT-4.1 despite the OpenAI finetuning API's safeguards. The finetuned model produces steganographic malicious outputs in response to hidden malicious prompts, while the user interface displays only a fully benign cover interaction. We also replicate the attack on three open-source models, Llama-3.3-70B-Instruct, Phi-4, and Mistral-Small-24B-Base-2501, confirming the generality of our method. We quantitatively evaluate our method on the AdvBench dataset, using Llama-Guard-3-8B for content safety classification. Across all four models, all stegotexts containing malicious content are incorrectly classified as safe.

Recently, NLP has seen a surge in the usage of large pre-trained models. Users download weights of models pre-trained on large datasets, then fine-tune the weights on a task of their choice. This raises the question of whether downloading untrusted pre-trained weights can pose a security threat. In this paper, we show that it is possible to construct ``weight poisoning'' attacks where pre-trained weights are injected with vulnerabilities that expose ``backdoors'' after fine-tuning, enabling the attacker to manipulate the model prediction simply by injecting an arbitrary keyword. We show that by applying a regularization method, which we call RIPPLe, and an initialization procedure, which we call Embedding Surgery, such attacks are possible even with limited knowledge of the dataset and fine-tuning procedure. Our experiments on sentiment classification, toxicity detection, and spam detection show that this attack is widely applicable and poses a serious threat. Finally, we outline practical defenses against such attacks. Code to reproduce our experiments is available at https://github.com/neulab/RIPPLe.

Adversarial attacks are carried out to reveal the vulnerability of deep neural networks. Textual adversarial attacking is challenging because text is discrete and a small perturbation can bring significant change to the original input. Word-level attacking, which can be regarded as a combinatorial optimization problem, is a well-studied class of textual attack methods. However, existing word-level attack models are far from perfect, largely because unsuitable search space reduction methods and inefficient optimization algorithms are employed. In this paper, we propose a novel attack model, which incorporates the sememe-based word substitution method and particle swarm optimization-based search algorithm to solve the two problems separately. We conduct exhaustive experiments to evaluate our attack model by attacking BiLSTM and BERT on three benchmark datasets. Experimental results demonstrate that our model consistently achieves much higher attack success rates and crafts more high-quality adversarial examples as compared to baseline methods. Also, further experiments show our model has higher transferability and can bring more robustness enhancement to victim models by adversarial training. All the code and data of this paper can be obtained on https://github.com/thunlp/SememePSO-Attack.

Membership inference attacks serves as useful tool for fair use of language models, such as detecting potential copyright infringement and auditing data leakage. However, many current state-of-the-art attacks require access to models' hidden states or probability distribution, which prevents investigation into more widely-used, API-access only models like GPT-4. In this work, we introduce N-Gram Coverage Attack, a membership inference attack that relies solely on text outputs from the target model, enabling attacks on completely black-box models. We leverage the observation that models are more likely to memorize and subsequently generate text patterns that were commonly observed in their training data. Specifically, to make a prediction on a candidate member, N-Gram Coverage Attack first obtains multiple model generations conditioned on a prefix of the candidate. It then uses n-gram overlap metrics to compute and aggregate the similarities of these outputs with the ground truth suffix; high similarities indicate likely membership. We first demonstrate on a diverse set of existing benchmarks that N-Gram Coverage Attack outperforms other black-box methods while also impressively achieving comparable or even better performance to state-of-the-art white-box attacks - despite having access to only text outputs. Interestingly, we find that the success rate of our method scales with the attack compute budget - as we increase the number of sequences generated from the target model conditioned on the prefix, attack performance tends to improve. Having verified the accuracy of our method, we use it to investigate previously unstudied closed OpenAI models on multiple domains. We find that more recent models, such as GPT-4o, exhibit increased robustness to membership inference, suggesting an evolving trend toward improved privacy protections.

Federated learning (FL) enables collaborative model training while preserving data privacy, but its decentralized nature exposes it to client-side data poisoning attacks (DPAs) and model poisoning attacks (MPAs) that degrade global model performance. While numerous proposed defenses claim substantial effectiveness, their evaluation is typically done in isolation with limited attack strategies, raising concerns about their validity. Additionally, existing studies overlook the mutual effectiveness of defenses against both DPAs and MPAs, causing fragmentation in this field. This paper aims to provide a unified benchmark and analysis of defenses against DPAs and MPAs, clarifying the distinction between these two similar but slightly distinct domains. We present a systematic taxonomy of poisoning attacks and defense strategies, outlining their design, strengths, and limitations. Then, a unified comparative evaluation across FL algorithms and data heterogeneity is conducted to validate their individual and mutual effectiveness and derive key insights for design principles and future research. Along with the analysis, we frame our work to a unified benchmark, FLPoison, with high modularity and scalability to evaluate 15 representative poisoning attacks and 17 defense strategies, facilitating future research in this domain. Code is available at https://github.com/vio1etus/FLPoison.

Recent developments in Large Language Models (LLMs) have manifested significant advancements. To facilitate safeguards against malicious exploitation, a body of research has concentrated on aligning LLMs with human preferences and inhibiting their generation of inappropriate content. Unfortunately, such alignments are often vulnerable: fine-tuning with a minimal amount of harmful data can easily unalign the target LLM. While being effective, such fine-tuning-based unalignment approaches also have their own limitations: (1) non-stealthiness, after fine-tuning, safety audits or red-teaming can easily expose the potential weaknesses of the unaligned models, thereby precluding their release/use. (2) non-persistence, the unaligned LLMs can be easily repaired through re-alignment, i.e., fine-tuning again with aligned data points. In this work, we show that it is possible to conduct stealthy and persistent unalignment on large language models via backdoor injections. We also provide a novel understanding on the relationship between the backdoor persistence and the activation pattern and further provide guidelines for potential trigger design. Through extensive experiments, we demonstrate that our proposed stealthy and persistent unalignment can successfully pass the safety evaluation while maintaining strong persistence against re-alignment defense.

Agent skills occupy a privileged position in the agent workflow, as agents are expected to implicitly follow and execute them, rendering third-party skills a vulnerable attack surface. Existing studies have revealed unsafe agent behaviors induced by skill-based attacks, but they primarily evaluate poisoned skills within a single task execution and enumerate harms through ad-hoc risk lists. To bridge these gaps, we introduce SkillHarm, a benchmark of skill-based attacks across the skill-use lifecycle, paired with a systematic taxonomy of skill-relevant risks. SkillHarm evaluates two attack scenarios: Fixed-Payload Poisoning (FPP), where a fixed poisoned skill package directly compromises any task session that invokes it, and Self-Mutating Poisoning (SMP), where an initially benign execution silently mutates persistent skill content, deferring harm until a subsequent reuse. It further defines 12 risk types based on the agent workflow component targeted by the harm: data pipelines, system environments, and agent autonomy. To instantiate these attacks at scale, we build AutoSkillHarm, an automated construction pipeline with coding agents driven by natural-language harnesses. The resulting benchmark contains 879 attack samples across 71 skills. Experiments show that current agents remain vulnerable with attack success rates up to 86.3% in FPP and 69.3% in SMP. Our analysis further reveals a latent risk: many apparent attack failures stem from the agent failing to engage with the poisoned file rather than genuine resistance, and current defenses still fail to reliably mitigate the threat.

Out-of-distribution (OOD) detection is critical to ensure the safe deployment of deep learning models in critical applications. Deep learning models can often misidentify OOD samples as in-distribution (ID) samples. This vulnerability worsens in the presence of spurious correlation in the training set. Likewise, in fine-grained classification settings, detection of fine-grained OOD samples becomes inherently challenging due to their high similarity to ID samples. However, current research on OOD detection has largely ignored these challenging scenarios, focusing instead on relatively easier (conventional) cases. In this work, we present a unified Approach to Spurious, fine-grained, and Conventional OOD Detection (ASCOOD). First, we propose synthesizing virtual outliers from ID data by approximating the destruction of invariant features. To this end, we identify invariant features with the pixel attribution method using the model being learned. This approach eliminates the burden of curating external OOD datasets. Then, we simultaneously incentivize ID classification and predictive uncertainty towards virtual outliers leveraging standardized feature representation. Our approach effectively mitigates the impact of spurious correlations and encourages capturing fine-grained attributes. Extensive experiments across seven datasets demonstrate the merit of ASCOOD in spurious, fine-grained, and conventional settings. The code is available at: https://github.com/sudarshanregmi/ASCOOD/

While prior research has proposed a plethora of methods that build neural classifiers robust against adversarial robustness, practitioners are still reluctant to adopt them due to their unacceptably severe clean accuracy penalties. This paper significantly alleviates this accuracy-robustness trade-off by mixing the output probabilities of a standard classifier and a robust classifier, where the standard network is optimized for clean accuracy and is not robust in general. We show that the robust base classifier's confidence difference for correct and incorrect examples is the key to this improvement. In addition to providing intuitions and empirical evidence, we theoretically certify the robustness of the mixed classifier under realistic assumptions. Furthermore, we adapt an adversarial input detector into a mixing network that adaptively adjusts the mixture of the two base models, further reducing the accuracy penalty of achieving robustness. The proposed flexible method, termed "adaptive smoothing", can work in conjunction with existing or even future methods that improve clean accuracy, robustness, or adversary detection. Our empirical evaluation considers strong attack methods, including AutoAttack and adaptive attack. On the CIFAR-100 dataset, our method achieves an 85.21% clean accuracy while maintaining a 38.72% ell_infty-AutoAttacked (epsilon = 8/255) accuracy, becoming the second most robust method on the RobustBench CIFAR-100 benchmark as of submission, while improving the clean accuracy by ten percentage points compared with all listed models. The code that implements our method is available at https://github.com/Bai-YT/AdaptiveSmoothing.

As a research community, we are still lacking a systematic understanding of the progress on adversarial robustness which often makes it hard to identify the most promising ideas in training robust models. A key challenge in benchmarking robustness is that its evaluation is often error-prone leading to robustness overestimation. Our goal is to establish a standardized benchmark of adversarial robustness, which as accurately as possible reflects the robustness of the considered models within a reasonable computational budget. To this end, we start by considering the image classification task and introduce restrictions (possibly loosened in the future) on the allowed models. We evaluate adversarial robustness with AutoAttack, an ensemble of white- and black-box attacks, which was recently shown in a large-scale study to improve almost all robustness evaluations compared to the original publications. To prevent overadaptation of new defenses to AutoAttack, we welcome external evaluations based on adaptive attacks, especially where AutoAttack flags a potential overestimation of robustness. Our leaderboard, hosted at https://robustbench.github.io/, contains evaluations of 120+ models and aims at reflecting the current state of the art in image classification on a set of well-defined tasks in ell_infty- and ell_2-threat models and on common corruptions, with possible extensions in the future. Additionally, we open-source the library https://github.com/RobustBench/robustbench that provides unified access to 80+ robust models to facilitate their downstream applications. Finally, based on the collected models, we analyze the impact of robustness on the performance on distribution shifts, calibration, out-of-distribution detection, fairness, privacy leakage, smoothness, and transferability.

As autonomous agents (e.g., OpenClaw) increasingly operate with deep system-level privileges to execute complex tasks, they introduce severe, unmitigated security risks. Current vulnerability analyses overwhelmingly focus on single-turn, stateless behaviors, overlooking the expanded attack surface inherent in stateful, multi-turn interactions and dynamic tool invocations. In this paper, we propose a novel, multi-dimensional evasion framework targeting LLM-based agent systems. We introduce three stealthy attack vectors: (1) Temporal evasion, which fragments malicious payloads across sequential interaction turns; (2) Spatial evasion, which conceals payloads within complex external artifacts that evade standard LLM parsing mechanisms; and (3) Semantic evasion, which obscures malicious intents beneath benign contextual noise. To systematically quantify these threats, we construct A3S-Bench, a comprehensive benchmark comprising 2,254 real-world agent execution trajectories. Evaluating a standard agent framework separately integrated with 10 mainstream LLM backbones against 20 practical threat scenarios, we demonstrate that our evasion framework elevates the average risk trigger rate from a 28.3\% baseline to 52.6\%. These findings reveal systemic, architecture-level vulnerabilities in current autonomous agent systems that existing defenses fail to address, highlighting an urgent need for defense mechanisms tailored to the unique threats.