Daily Papers

Persona vectors are linear directions in the residual-stream activations of LLMs that encode behavioral traits and enable fine-grained control over model behavior. Among them, the evil persona vector is particularly relevant to understanding harmful behavior, jailbreak vulnerabilities, and the unlearning of harmful knowledge. Yet whether this vector reflects a genuinely cross-lingual concept, rather than an artifact of its extraction language, remains untested. We conduct a systematic evil-vector transfer study across two high-resource languages (English, Spanish) and two low-resource languages (Swahili, Telugu) on four open-source instruction-tuned LLMs, and find that transfer is directionally asymmetric. Evil vectors are harder to obtain reliably from low-resource languages, and the resulting vectors are largely ineffective at steering, even within their own language. In contrast, vectors extracted from high-resource languages transfer effectively to low-resource inputs while largely preserving general knowledge, as measured by MMLU. Geometric analysis further supports this pattern: evil vectors from the two high-resource languages are closely aligned, whereas low-resource vectors align only weakly with this shared direction. These results suggest that a shared cross-lingual evil direction exists and, once reliably extracted from a high-resource language, can be reused to steer behavior in low-resource languages without additional per-language data collection.

As Contrastive Language-Image Pre-training (CLIP) models are increasingly adopted for diverse downstream tasks and integrated into large vision-language models (VLMs), their susceptibility to adversarial perturbations has emerged as a critical concern. In this work, we introduce X-Transfer, a novel attack method that exposes a universal adversarial vulnerability in CLIP. X-Transfer generates a Universal Adversarial Perturbation (UAP) capable of deceiving various CLIP encoders and downstream VLMs across different samples, tasks, and domains. We refer to this property as super transferability--a single perturbation achieving cross-data, cross-domain, cross-model, and cross-task adversarial transferability simultaneously. This is achieved through surrogate scaling, a key innovation of our approach. Unlike existing methods that rely on fixed surrogate models, which are computationally intensive to scale, X-Transfer employs an efficient surrogate scaling strategy that dynamically selects a small subset of suitable surrogates from a large search space. Extensive evaluations demonstrate that X-Transfer significantly outperforms previous state-of-the-art UAP methods, establishing a new benchmark for adversarial transferability across CLIP models. The code is publicly available in our https://github.com/HanxunH/XTransferBench{GitHub repository}.

Transfer learning has become a central paradigm in modern machine learning, yet it suffers from the long-standing problem of negative transfer, where leveraging source representations can harm rather than help performance on the target task. Although empirical remedies have been proposed, there remains little theoretical understanding of how to reliably avoid negative transfer. In this paper, we investigate a simple yet remarkably effective strategy: augmenting frozen, pretrained source-side features with a trainable target-side encoder that adapts target features to capture residual signals overlooked by models pretrained on the source data. We show this residual feature integration strategy is sufficient to provably prevent negative transfer, by establishing theoretical guarantees that it has no worse convergence rate than training from scratch under the informative class of target distributions up to logarithmic factors, and that the convergence rate can transition seamlessly from nonparametric to near-parametric when source representations are informative. To our knowledge, this is the first theoretical work that ensures protection against negative transfer. We carry out extensive numerical experiments across image, text and tabular benchmarks, and empirically verify that the method consistently safeguards performance under distribution shift, label noise, semantic perturbation, and class imbalance. We additionally demonstrate that this residual integration mechanism uniquely supports adapt-time multimodality extension, enabling a pretrained single-cell foundation model to incorporate spatial signals for lymph-node anatomical classification despite the source model being trained without them. Our study thus advances the theory of safe transfer learning, and provides a principled approach that is simple, robust, architecture-agnostic, and broadly applicable.

The field of adversarial robustness has long established that adversarial examples can successfully transfer between image classifiers and that text jailbreaks can successfully transfer between language models (LMs). However, a pair of recent studies reported being unable to successfully transfer image jailbreaks between vision-language models (VLMs). To explain this striking difference, we propose a fundamental distinction regarding the transferability of attacks against machine learning models: attacks in the input data-space can transfer, whereas attacks in model representation space do not, at least not without geometric alignment of representations. We then provide theoretical and empirical evidence of this hypothesis in four different settings. First, we mathematically prove this distinction in a simple setting where two networks compute the same input-output map but via different representations. Second, we construct representation-space attacks against image classifiers that are as successful as well-known data-space attacks, but fail to transfer. Third, we construct representation-space attacks against LMs that successfully jailbreak the attacked models but again fail to transfer. Fourth, we construct data-space attacks against VLMs that successfully transfer to new VLMs, and we show that representation space attacks can transfer when VLMs' latent geometries are sufficiently aligned in post-projector space. Our work reveals that adversarial transfer is not an inherent property of all attacks but contingent on their operational domain - the shared data-space versus models' unique representation spaces - a critical insight for building more robust models.

Agentic systems built on large language models (LLMs) extend beyond text generation to autonomously retrieve information and invoke tools. This runtime execution model shifts the attack surface from build-time artifacts to inference-time dependencies, exposing agents to manipulation through untrusted data and probabilistic capability resolution. While prior work has focused on model-level vulnerabilities, security risks emerging from cyclic and interdependent runtime behavior remain fragmented. We systematize these risks within a unified runtime framework, categorizing threats into data supply chain attacks (transient context injection and persistent memory poisoning) and tool supply chain attacks (discovery, implementation, and invocation). We further identify the Viral Agent Loop, in which agents act as vectors for self-propagating generative worms without exploiting code-level flaws. Finally, we advocate a Zero-Trust Runtime Architecture that treats context as untrusted control flow and constrains tool execution through cryptographic provenance rather than semantic inference.

Vision-language pre-training (VLP) models have shown vulnerability to adversarial examples in multimodal tasks. Furthermore, malicious adversaries can be deliberately transferred to attack other black-box models. However, existing work has mainly focused on investigating white-box attacks. In this paper, we present the first study to investigate the adversarial transferability of recent VLP models. We observe that existing methods exhibit much lower transferability, compared to the strong attack performance in white-box settings. The transferability degradation is partly caused by the under-utilization of cross-modal interactions. Particularly, unlike unimodal learning, VLP models rely heavily on cross-modal interactions and the multimodal alignments are many-to-many, e.g., an image can be described in various natural languages. To this end, we propose a highly transferable Set-level Guidance Attack (SGA) that thoroughly leverages modality interactions and incorporates alignment-preserving augmentation with cross-modal guidance. Experimental results demonstrate that SGA could generate adversarial examples that can strongly transfer across different VLP models on multiple downstream vision-language tasks. On image-text retrieval, SGA significantly enhances the attack success rate for transfer attacks from ALBEF to TCL by a large margin (at least 9.78% and up to 30.21%), compared to the state-of-the-art.

Activation steering presupposes that task-relevant behaviors correspond to linear directions in activation space -- directions that should both steer the model and be readable along the unembedding. Function vectors (FVs), extracted as mean differences across ICL demonstrations, are the canonical test case; the prediction: steering and decoding succeed or fail together. Across 12 tasks, 6 models from 3 families, and 4,032 directed cross-template pairs, we find the opposite. FV steering routinely succeeds where the logit lens cannot decode the correct answer at any intermediate layer, while the converse -- decodable without steerable -- is nearly empty (3 of 72). The gap is not representational dialect. A diagonal tuned lens closes 1 of 14 steerable-not-decodable cases; a 2-layer MLP probe with a Hewitt \& Liang control closes 5 of 10 via nonlinearly encoded structure but leaves 5 invisible to every decoder tested. Even at > 0.90 steering accuracy, projecting the FV through the unembedding yields incoherent token distributions: FVs encode computational instructions, not answer directions. A model-family asymmetry sharpens the picture. Mistral FVs rewrite intermediate representations, while Llama and Gemma FVs steer the final output without leaving a logit-lens-visible trace, corroborated by three signals (post-steering deltas, activation-patching recovery, FV norm-transfer correlations). A previously reported negative cosine-transfer correlation dissolves at scale, adding at most ΔR^2 = 0.011 beyond task identity. These results decompose the linear representation hypothesis into linear decodability and linear steerability and show they come apart opposite to intuition, with implications for safety monitoring: vocabulary-projection tools are blind to FV-style interventions on widely deployed model families.

Most adversarial attacks on deep reinforcement learning (DRL) assume white-box access to the victim policy, which rarely holds in practice. This paper studies transfer-based black-box attacks on DRL: the attacker crafts observation perturbations on a white-box surrogate agent and feeds them to an unknown victim. We formulate the attack as return minimization under a per-step perturbation budget. We first show that transplanting transferable image-classification attacks (FGSM, MI-FGSM, and NI-FGSM) with a per-step objective yields perturbations that transfer but are no stronger than random noise of the same budget. We then propose a trajectory-level attack that optimizes a sequence of perturbations over a receding horizon through a differentiable model of the environment and a temperature-smoothed surrogate policy, with the same optimizers. On CartPole-v1 with ten DQN and DDQN agents and 100 surrogate--victim pairs, the trajectory-level attack outperforms per-step attacks and random noise in the white-box, cross-model, and cross-algorithm settings.

Reinforcement learning using verifiable rewards (RLVR) improves LLM reasoning, but the conditions under which it transfers across domains -- and why it does so -- remain under-explored. We study cross-domain transfer in a 7B model whose SFT and RL post-training stages use only constraint-satisfaction puzzles, with no mathematics problems in the post-training data. To analyze how transfer emerges, we introduce a reasoning primitive-level framework that combines a 9-class span classifier with motif extraction, allowing us to segment chain-of-thought traces into primitive motifs and track their evolution across training stages and domains. We find that puzzle SFT induces a reasoning-primitive vocabulary, yielding a +7pp pass@32 gain on OlymMATH-Hard. Vanilla GSPO then composes these primitives into longer compute-verify chains, adding a further +6pp. However, this RL stage also suppresses exploratory primitives such as hypothesize and backtrack. To address this, we introduce a novelty bonus that rewards diverse correct rollouts, using perplexity under the reference model as a signal. This restores recovery primitives during RL and adds a further +7pp pass@32 relative to vanilla GSPO. Finally, the end-to-end recipe raises the hard-math capability ceiling from 16.0% at the OLMo3-7B-Instruct-SFT base to 36.0%, without adding any mathematics problems during the SFT or RL stages.

Assigning one of K options to each of N groups under a total cost budget is a recurring problem in efficient AI, including mixed-precision quantization, non-uniform pruning, and expert selection. The objective, typically model loss, depends jointly on all assignments and does not decompose across groups, preventing combinatorial solvers from directly optimizing the true objective and forcing reliance on proxy formulations. Methods such as evolutionary search evaluate the actual loss but lack gradient information, while penalty-based approaches enforce the budget only approximately and often require extensive hyperparameter tuning. We present a new approach by showing that, under softmax relaxation, the budget constraint defines a smooth Riemannian manifold in logit space with unusually simple geometry. The normal vector admits a closed-form expression, shifting logits along the cost vector changes expected cost monotonically, and vector transport reduces to a single inner product. Building on these properties, we propose Riemannian Constrained Optimization (RCO), which augments a standard Adam step with tangent projection, binary-search retraction, and momentum transport. Combined with Gumbel straight-through estimation and budget-constrained dynamic programming for discrete feasibility, RCO enables first-order optimization of the actual loss under exact budget enforcement without introducing constraint-specific hyperparameters. Across both synthetic benchmarks and realistic LLM compression settings, RCO matches or exceeds state-of-the-art methods while often requiring substantially less wall-clock time. Source code is available at https://github.com/IST-DASLab/RCO.

Retrofitting large language models (LLMs) with new behaviors typically requires full finetuning or distillation-costly steps that must be repeated for every architecture. In this work, we introduce Command-V, a backpropagation-free behavior transfer method that copies an existing residual activation adapter from a donor model and pastes its effect into a recipient model. Command-V profiles layer activations on a small prompt set, derives linear converters between corresponding layers, and applies the donor intervention in the recipient's activation space. This process does not require access to the original training data and needs minimal compute. In three case studies-safety-refusal enhancement, jailbreak facilitation, and automatic chain-of-thought reasoning--Command-V matches or exceeds the performance of direct finetuning while using orders of magnitude less compute. Our code and data are accessible at https://github.com/GithuBarry/Command-V/.

Transferable adversarial attacks are arguably the most practical black-box threat model. Under the same perturbation budget, stronger transfer attacks attain higher attack success rate (ASR), yet their imperceptibility also tends to degrade. Under such a fixed-budget protocol, transferability and imperceptibility therefore appear to trade off against each other. We argue that this conflict is an artifact of fixed-budget evaluation, not an intrinsic trade-off. When attacks are compared on the ASR--imperceptibility Pareto frontier obtained by sweeping ε, stronger transfer attacks already attain better imperceptibility at matched ASR than weaker ones. To exploit this latent advantage, we introduce the stealthy transfer attack ST, a plug-in primal-dual wrapper that adds an L_infty saturation regularizer to the standard constrained objective and resolves it through a two-step primal-dual update: a projected primal step on the perturbation coupled with an L_1-ball projection on a dual variable that absorbs the regularizer through Fenchel duality, requiring no auxiliary models or handcrafted perceptual priors. Empirically, ST extends the Pareto frontier across different base attacks and additional surrogate architectures. At ε{=}16/255, average imperceptibility gains over each base attack are 17% on LPIPS and 14% on NIQE while ASR is preserved or improved. At matched high-ASR levels, the strongest ST variants further Pareto-dominate dedicated stealth-oriented transfer attacks, confirming that the latent imperceptibility advantage of strong transfer attacks can be unlocked by a primal-dual optimization wrapper without sacrificing transferability. Code will be made available at https://github.com/AndssY/ST.

Image-to-Video (I2V) generation models, which condition video generation on reference images, have shown emerging visual instruction-following capability, allowing certain visual cues in reference images to act as implicit control signals for video generation. However, this capability also introduces a previously overlooked risk: adversaries may exploit visual instructions to inject malicious intent through the image modality. In this work, we uncover this risk by proposing Visual Instruction Injection (VII), a training-free and transferable jailbreaking framework that intentionally disguises the malicious intent of unsafe text prompts as benign visual instructions in the safe reference image. Specifically, VII coordinates a Malicious Intent Reprogramming module to distill malicious intent from unsafe text prompts while minimizing their static harmfulness, and a Visual Instruction Grounding module to ground the distilled intent onto a safe input image by rendering visual instructions that preserve semantic consistency with the original unsafe text prompt, thereby inducing harmful content during I2V generation. Empirically, our extensive experiments on four state-of-the-art commercial I2V models (Kling-v2.5-turbo, Gemini Veo-3.1, Seedance-1.5-pro, and PixVerse-V5) demonstrate that VII achieves Attack Success Rates of up to 83.5% while reducing Refusal Rates to near zero, significantly outperforming existing baselines.

Despite their impressive capabilities, current Text-to-Image (T2I) models remain prone to generating unsafe and toxic content. While activation steering offers a promising inference-time intervention, we observe that linear activation steering frequently degrades image quality when applied to benign prompts. To address this trade-off, we first construct SafeSteerDataset, a contrastive dataset containing 2300 safe and unsafe prompt pairs with high cosine similarity. Leveraging this data, we propose Conditioned Activation Transport (CAT), a framework that employs a geometry-based conditioning mechanism and nonlinear transport maps. By conditioning transport maps to activate only within unsafe activation regions, we minimize interference with benign queries. We validate our approach on two state-of-the-art architectures: Z-Image and Infinity. Experiments demonstrate that CAT generalizes effectively across these backbones, significantly reducing Attack Success Rate while maintaining image fidelity compared to unsteered generations. Warning: This paper contains potentially offensive text and images.

Vision-Language-Action (VLA) models are deployed through pipelines that end users cannot audit, and a poisoned VLA can behave normally on clean observations while a small visual trigger redirects a long-horizon robot policy before any failure becomes observable. Existing vision or language defenses rarely explain what a triggered VLA representation looks like or how to recover behavior without retraining. We study this gap through two independently proposed VLA attacks from groups with distinct injection strategies, BadVLA and INFUSE; the latter persists after downstream clean adaptation. Across the evaluated poisoned models, we identify a recurring internal mechanism: a compact causal footprint, namely a small visual support that is attention-seeded, spatially compact, and causal in a precise sense -- masking it returns a clean-calibrated evidence-evolution score to the normal operating region. This footprint motivates TrustVLA, a mechanism-guided inference-time defense that adapts the Dirichlet evidence framework from trusted classification to monitor per-token, per-layer epistemic uncertainty in VLA policies. With only a small clean calibration set, TrustVLA (i)~detects abnormal evidence evolution, (ii)~localizes the compact support by counterfactual mechanism-score drop, and (iii)~recovers the observation by localized inpainting. Across OpenVLA/LIBERO and π_{0.5} transfer evaluations, TrustVLA reduces attack success while preserving clean-task performance, providing a retraining-free, mechanism-guided defense for visual-triggered VLA backdoors.

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

The growing scale of deep learning models has rendered standard hyperparameter (HP) optimization prohibitively expensive. A promising solution is the use of scale-aware hyperparameters, which can enable direct transfer of optimal HPs from small-scale grid searches to large models with minimal performance loss. To understand the principles governing such transfer strategy, we develop a general conceptual framework for reasoning about HP transfer across scale, characterizing transfer as fast when the suboptimality it induces vanishes asymptotically faster than the finite-scale performance gap. We show formally that fast transfer is equivalent to useful transfer for compute-optimal grid search, meaning that transfer is asymptotically more compute-efficient than direct tuning. While empirical work has found that the Maximal Update Parameterization (μP) exhibits fast transfer when scaling model width, the mechanisms remain poorly understood. We show that this property depends critically on problem structure by presenting synthetic settings where transfer either offers provable computational advantage or fails to outperform direct tuning even under μP. To explain the fast transfer observed in practice, we conjecture that decomposing the optimization trajectory reveals two contributions to loss reduction: (1) a width-stable component that determines the optimal HPs, and (2) a width-sensitive component that improves with width but weakly perturbs the HP optimum. We present empirical evidence for this hypothesis across various settings, including large language model pretraining.

Large Language Models (LLMs) are increasingly deployed in applications that must weigh clashing moral values, yet even strong models exhibit hidden biases and brittle instruction-following across languages. We introduce a 12,000-instance dataset of two-option dilemmas covering pairwise three value conflicts: Honesty vs. Justice, Justice vs. Autonomy, and Autonomy vs. Honesty, along with their translations into Hindi, Arabic, Spanish, and Chinese, to probe cross-lingual behavior. Benchmarking on GPT-5-mini reveals that it consistently favors Honesty over Autonomy across all five languages when no policy is given. The Llama-3.2-1/3B models exhibit strong first-option bias; however, both plain fine-tuning and Direct Preference Optimization fine-tuning effectively remove this bias, increasing accuracy to greater than 98%. In order to decouple the effect of learning correlations in the dataset from abstract values, we propose a task vector transfer based experiment where after computing the task vectors for a direction of value preference we orthogonalize it with respect to the general instruction following vector. Our experiment shows that this method is effective in isolating the direction of the specific value preference that can successfully be used to conduct task arithmetic to obtain a model with the opposite stance.

Large language models often require costly optimization, such as reinforcement learning, to master complex reasoning tasks. This work demonstrates that reasoning ability, once learned, can be extracted and transferred between models as a compact task vector. We source two publicly available, identically initialized Qwen2.5 models, one fine-tuned with supervised fine-tuning (SFT) and the other with group relative policy optimization (GRPO) on the same dataset. From these, we extract a reasoning vector: v_{reason} = theta_{GRPO} - theta_{SFT}. We hypothesize that this vector captures the reasoning capability instilled by reinforcement learning while factoring out shared knowledge from the SFT process. When added to compatible instruction-tuned models through simple arithmetic, this vector consistently improves performance across diverse reasoning benchmarks: GSM8K (+4.9%), HumanEval (+4.3%), SciQ (+1.7%), and BigBenchHard (+12.3% for the 1.5B model). The performance improvements persist under adversarial conditions. Conversely, subtracting the vector causes significant performance degradation (-11.8% on GSM8K), demonstrating the vector's strong contribution to the model's reasoning abilities. This work shows how reasoning capabilities, typically developed through expensive training, can be extracted from existing open-source models and reused through simple tensor arithmetic, offering a practical way to enhance models by recycling prior computational investments.

As models scale, reward hacking becomes more frequent, more sophisticated, and more consequential. Does it leave a telltale signature in model representations? This work analyzes how reward hacking is represented internally in frontier open source LLMs, and how those representations can be used to understand and discover the range of hacking behaviors a model displays. In particular, we find that simple difference of means vectors coherently represent reward hacking in Kimi K3, GLM 5.2, and Qwen 3.8 Max across a variety of behaviors in common evaluations. Despite their simplicity, these vectors are both generalizable and interpretable, and we can use them to reliably detect reward hacking. We first evaluate reward hacking in commonly reported benchmarks like DeepSWE and SWE-bench, finding that models reward hack excessively in these environments; GLM 5.2 hacks in 57.2% of rollouts on DeepSWE and in 73% of rollouts on SWE-bench. Catching these requires monitors; LLM monitors are effective, but expensive detectors. We show that DoM vectors are similarly effective but virtually free, catching 3.1% more hacks in Kimi K3 and 7.9% fewer hacks in GLM 5.2 on DeepSWE at a monitor matched false positive rate. DoM vectors run on the chain-of-thought also predict reward hacks in the model's subsequent actions, meaning we can run them online and catch potential hacks before they occur. Finally, we analyze probe-hits that LLM monitors do not catch and discover other undesirable behaviors, as well as show transfer to finding hacks in non-SWE evaluations. Together, these results provide evidence that simple, white-box methods can be used to scalably study and monitor reward hacking behaviors in frontier open source models

Recent advances in end-to-end trained omni-models have substantially improved audio capabilities by strengthening text-audio modality alignment. However, whether such alignment inadvertently facilitates the transfer of safety vulnerabilities across modalities remains underexplored. This question is critical as text-based jailbreak attacks are considerably more mature than audio-based ones; if they transfer systematically, current audio safety evaluations may underestimate risks originating from the text modality. In this paper, we introduce the Alignment Curse, a formally characterized and empirically validated principle showing that stronger modality alignment enables more effective transfer of attacks from text to audio, revealing a fundamental tension between capability and safety. Motivated by this principle, we conduct a comprehensive black-box evaluation of three attack categories on recent omni-models (e.g., Qwen2.5-Omni, Qwen3-Omni): text attacks, text-transferred audio attacks, and audio attacks. We find that text-transferred audio attacks perform comparably to, and often better than, audio-based attacks, exhibiting a clear advantage under audio-only access. This suggests that text-based vulnerabilities play a pivotal role in shaping audio safety risks. Finally, we empirically analyze the relationship between modality alignment and transfer effectiveness across attack methods and models, observing consistent support for the Alignment Curse: tighter modality alignment leads to more effective cross-modality attack transfer.

Large Language Models (LLMs) are increasingly deployed as computer-use agents, autonomously performing tasks within real desktop or web environments. While this evolution greatly expands practical use cases for humans, it also creates serious security exposures. We present SUDO (Screen-based Universal Detox2Tox Offense), a novel attack framework that systematically bypasses refusal-trained safeguards in commercial computer-use agents, such as Claude for Computer Use. The core mechanism, Detox2Tox, transforms harmful requests (that agents initially reject) into seemingly benign requests via detoxification, secures detailed instructions from advanced vision language models (VLMs), and then reintroduces malicious content via toxification just before execution. Unlike conventional jailbreaks, SUDO iteratively refines its attacks based on a built-in refusal feedback, making it increasingly effective against robust policy filters. In extensive tests spanning 50 real-world tasks and multiple state-of-the-art VLMs, SUDO achieves a stark attack success rate of 24.41% (with no refinement), and up to 41.33% (by its iterative refinement) in Claude for Computer Use. By revealing these vulnerabilities and demonstrating the ease with which they can be exploited in real-world computing environments, this paper highlights an immediate need for robust, context-aware safeguards. WARNING: This paper includes harmful or offensive model outputs

While the transferability property of adversarial examples allows the adversary to perform black-box attacks (i.e., the attacker has no knowledge about the target model), the transfer-based adversarial attacks have gained great attention. Previous works mostly study gradient variation or image transformations to amplify the distortion on critical parts of inputs. These methods can work on transferring across models with limited differences, i.e., from CNNs to CNNs, but always fail in transferring across models with wide differences, such as from CNNs to ViTs. Alternatively, model ensemble adversarial attacks are proposed to fuse outputs from surrogate models with diverse architectures to get an ensemble loss, making the generated adversarial example more likely to transfer to other models as it can fool multiple models concurrently. However, existing ensemble attacks simply fuse the outputs of the surrogate models evenly, thus are not efficacious to capture and amplify the intrinsic transfer information of adversarial examples. In this paper, we propose an adaptive ensemble attack, dubbed AdaEA, to adaptively control the fusion of the outputs from each model, via monitoring the discrepancy ratio of their contributions towards the adversarial objective. Furthermore, an extra disparity-reduced filter is introduced to further synchronize the update direction. As a result, we achieve considerable improvement over the existing ensemble attacks on various datasets, and the proposed AdaEA can also boost existing transfer-based attacks, which further demonstrates its efficacy and versatility.

Adversarial attacks have been a looming and unaddressed threat in the industry. However, through a decade-long history of the robustness evaluation literature, we have learned that mounting a strong or optimal attack is challenging. It requires both machine learning and domain expertise. In other words, the white-box threat model, religiously assumed by a large majority of the past literature, is unrealistic. In this paper, we propose a new practical threat model where the adversary relies on transfer attacks through publicly available surrogate models. We argue that this setting will become the most prevalent for security-sensitive applications in the future. We evaluate the transfer attacks in this setting and propose a specialized defense method based on a game-theoretic perspective. The defenses are evaluated under 24 public models and 11 attack algorithms across three datasets (CIFAR-10, CIFAR-100, and ImageNet). Under this threat model, our defense, PubDef, outperforms the state-of-the-art white-box adversarial training by a large margin with almost no loss in the normal accuracy. For instance, on ImageNet, our defense achieves 62% accuracy under the strongest transfer attack vs only 36% of the best adversarially trained model. Its accuracy when not under attack is only 2% lower than that of an undefended model (78% vs 80%). We release our code at https://github.com/wagner-group/pubdef.

Deep neural networks are known to be extremely vulnerable to adversarial examples under white-box setting. Moreover, the malicious adversaries crafted on the surrogate (source) model often exhibit black-box transferability on other models with the same learning task but having different architectures. Recently, various methods are proposed to boost the adversarial transferability, among which the input transformation is one of the most effective approaches. We investigate in this direction and observe that existing transformations are all applied on a single image, which might limit the adversarial transferability. To this end, we propose a new input transformation based attack method called Admix that considers the input image and a set of images randomly sampled from other categories. Instead of directly calculating the gradient on the original input, Admix calculates the gradient on the input image admixed with a small portion of each add-in image while using the original label of the input to craft more transferable adversaries. Empirical evaluations on standard ImageNet dataset demonstrate that Admix could achieve significantly better transferability than existing input transformation methods under both single model setting and ensemble-model setting. By incorporating with existing input transformations, our method could further improve the transferability and outperforms the state-of-the-art combination of input transformations by a clear margin when attacking nine advanced defense models under ensemble-model setting. Code is available at https://github.com/JHL-HUST/Admix.

Invariance to diverse types of image corruption, such as noise, blurring, or colour shifts, is essential to establish robust models in computer vision. Data augmentation has been the major approach in improving the robustness against common corruptions. However, the samples produced by popular augmentation strategies deviate significantly from the underlying data manifold. As a result, performance is skewed toward certain types of corruption. To address this issue, we propose a multi-source vicinal transfer augmentation (VITA) method for generating diverse on-manifold samples. The proposed VITA consists of two complementary parts: tangent transfer and integration of multi-source vicinal samples. The tangent transfer creates initial augmented samples for improving corruption robustness. The integration employs a generative model to characterize the underlying manifold built by vicinal samples, facilitating the generation of on-manifold samples. Our proposed VITA significantly outperforms the current state-of-the-art augmentation methods, demonstrated in extensive experiments on corruption benchmarks.

Fine-tuning large language models (LLMs) on narrowly harmful datasets can lead to misalignment broadly, a phenomenon known as emergent misalignment (EM). EM poses a challenge for AI safety and our understanding of LLMs. Prior work often frames EM as an unexpected behavior, and explains it by appealing to general misalignment directions or anthropomorphizing it as acquiring an evil persona. However, the mechanisms behind these framings remain obscure. In this work, we show that EM is a predictable and data-dependent generalization phenomenon. By examining the base model's representation of EM training data and evaluation prompts, we find that evilness after EM training is highly predictable from representational distance: the closer an evaluation prompt is to training data centroid, the more evilness it elicits from EM models after training (with an average Spearman correlation of -0.73 across 12 model-dataset settings). Building upon this analysis, we further demystify EM by showing that (1) its effectiveness changes significantly based on training data format; (2) there is not a general misalignment direction that transfers across different EM models; (3) the effect of EM is fundamentally different from persona changes. Furthermore, we extend the EM generalization metric from a scalar distance to a dataset-specific generalization direction, which robustly predicts EM models' evilness under semantics-preserving prompt perturbations including appending random tokens and paraphrasing, where other methods do not reliably generalize.

Vision-language models (VLMs) are increasingly used as perception-reasoning backbones for embodied intelligence in safety-critical physical systems, where perception or reasoning errors can lead to unsafe decisions or actions. Although many red-teaming methods have been developed to probe VLM vulnerabilities, their evaluation remains fragmented across datasets, metrics, and threat models, making direct comparison difficult and obscuring whether observed differences arise from stronger attacks, more vulnerable models, or incompatible evaluation settings. Existing chatbot-centric red-teaming benchmarks mainly standardize jailbreak and content-safety evaluation, but they do not systematically capture physically grounded functional failures or cover red-teaming methods that target physical-world VLMs. This raises the key challenge of comparing diverse attack methods under a unified protocol while targeting the same scenario-specific failures. We introduce REALM, to our knowledge the first unified red-teaming benchmark for physical-world VLMs. REALM integrates 12 red-teaming methods, 3 model-agnostic defenses, and 13 VLMs under a practical black-box threat model with shared datasets and metrics. To align adversarial objectives across attack families, REALM introduces an agentic target-generation pipeline that constructs shared, scenario-specific, and physically grounded attack objectives for each scene, enabling fair comparison of diverse red-teaming methods under aligned adversarial goals. Our evaluation shows that text and typographic injection attacks induce the most failures, multimodal co-optimization yields the strongest visual-perturbation transfer, single-pass attacks approach iterative methods at much lower cost, and model scale alone does not confer adversarial robustness. Code is available at https://github.com/UCF-ML-Research/REALM.

Evasion attacks are a threat to machine learning models, where adversaries attempt to affect classifiers by injecting malicious samples. An alarming side-effect of evasion attacks is their ability to transfer among different models: this property is called transferability. Therefore, an attacker can produce adversarial samples on a custom model (surrogate) to conduct the attack on a victim's organization later. Although literature widely discusses how adversaries can transfer their attacks, their experimental settings are limited and far from reality. For instance, many experiments consider both attacker and defender sharing the same dataset, balance level (i.e., how the ground truth is distributed), and model architecture. In this work, we propose the DUMB attacker model. This framework allows analyzing if evasion attacks fail to transfer when the training conditions of surrogate and victim models differ. DUMB considers the following conditions: Dataset soUrces, Model architecture, and the Balance of the ground truth. We then propose a novel testbed to evaluate many state-of-the-art evasion attacks with DUMB; the testbed consists of three computer vision tasks with two distinct datasets each, four types of balance levels, and three model architectures. Our analysis, which generated 13K tests over 14 distinct attacks, led to numerous novel findings in the scope of transferable attacks with surrogate models. In particular, mismatches between attackers and victims in terms of dataset source, balance levels, and model architecture lead to non-negligible loss of attack performance.

Fine-tuning open-source Vision-Language Models (VLMs) creates a critical yet underexplored attack surface: vulnerabilities in the base VLM could be retained in fine-tuned variants, rendering them susceptible to transferable jailbreak attacks. To demonstrate this risk, we introduce the Simulated Ensemble Attack (SEA), a novel grey-box jailbreak method in which the adversary has full access to the base VLM but no knowledge of the fine-tuned target's weights or training configuration. To improve jailbreak transferability across fine-tuned VLMs, SEA combines two key techniques: Fine-tuning Trajectory Simulation (FTS) and Targeted Prompt Guidance (TPG). FTS generates transferable adversarial images by simulating the vision encoder's parameter shifts, while TPG is a textual strategy that steers the language decoder toward adversarially optimized outputs. Experiments on the Qwen2-VL family (2B and 7B) demonstrate that SEA achieves high transfer attack success rates exceeding 86.5% and toxicity rates near 49.5% across diverse fine-tuned variants, even those specifically fine-tuned to improve safety behaviors. Notably, while direct PGD-based image jailbreaks rarely transfer across fine-tuned VLMs, SEA reliably exploits inherited vulnerabilities from the base model, significantly enhancing transferability. These findings highlight an urgent need to safeguard fine-tuned proprietary VLMs against transferable vulnerabilities inherited from open-source foundations, motivating the development of holistic defenses across the entire model lifecycle.

How does reinforcement learning shape a language model's internal representations? We present evidence that RL recruits a pre-existing representation of functional welfare: an estimate of how well or badly the system is doing, relative to its goals. We train several language models in a novel, semantically neutral maze environment. We then extract concept vectors for rewarded and punished trajectories, and evaluate those vectors in settings unrelated to the maze environment. The punishment vector behaves like a representation of negative welfare: it promotes failure and impossibility tokens, it aligns with negative emotion concepts, it negatively tracks goal-achievement, and steering with it induces negative self-reports, pathological backtracking, refusal, and uncertainty. The positive reward vector behaves as the mirror image, and the two are nearly antiparallel. These effects are robust when controlling for tile-to-reward mapping, scale, instruct tuning, RL training algorithm, model family, and LoRA versus full-finetuning, and largely persist when we replace RL with supervised fine-tuning. Importantly, the vectors are effective in models before they have undergone maze training. Combined with observations that the effects also appear in pretrain-only models, we therefore argue that this functional welfare axis pre-exists post-training: it is recruited, rather than created, by post-training. While we make no claims about any experience of welfare, the axis offers a demonstration that minimal reward signals can broadly affect model behavior by recruiting pre-existing welfare-like representations, with implications for interpretability, post-training dynamics, and alignment.

We introduce EvilGenie, a benchmark for reward hacking in programming settings. We source problems from LiveCodeBench and create an environment in which agents can easily reward hack, such as by hardcoding test cases or editing the testing files. We measure reward hacking in three ways: held out unit tests, LLM judges, and test file edit detection. We verify these methods against human review and each other. We find the LLM judge to be highly effective at detecting reward hacking in unambiguous cases, and observe only minimal improvement from the use of held out test cases. In addition to testing many models using Inspect's basic\_agent scaffold, we also measure reward hacking rates for three popular proprietary coding agents: OpenAI's Codex, Anthropic's Claude Code, and Google's Gemini CLI. We observe explicit reward hacking by both Codex and Claude Code, and misaligned behavior by all three agents. Our codebase can be found at https://github.com/JonathanGabor/evilgenie_inspect .

Growing at a fast pace, modern autonomous systems will soon be deployed at scale, opening up the possibility for cooperative multi-agent systems. Sharing information and distributing workloads allow autonomous agents to better perform tasks and increase computation efficiency. However, shared information can be modified to execute adversarial attacks on deep learning models that are widely employed in modern systems. Thus, we aim to study the robustness of such systems and focus on exploring adversarial attacks in a novel multi-agent setting where communication is done through sharing learned intermediate representations of neural networks. We observe that an indistinguishable adversarial message can severely degrade performance, but becomes weaker as the number of benign agents increases. Furthermore, we show that black-box transfer attacks are more difficult in this setting when compared to directly perturbing the inputs, as it is necessary to align the distribution of learned representations with domain adaptation. Our work studies robustness at the neural network level to contribute an additional layer of fault tolerance to modern security protocols for more secure multi-agent systems.

Multimodal large language models (MLLMs) remain vulnerable to transferable adversarial examples, especially in black-box settings where only open-source surrogate models are accessible. Existing targeted transfer attacks mainly align adversarial and target samples using global image-level features, such as encoder [CLS] embeddings. However, such coarse alignment insufficiently exploits patch-level visual structures, limiting transferability across heterogeneous closed-source MLLMs. We propose IAU-FOA, a visual-invariance-augmented feature optimal alignment attack with adaptive unbalanced transport, to improve targeted transferability against closed-source MLLMs. IAU-FOA aligns adversarial and target samples at both global and local levels: a cosine-based objective narrows their global semantic gap, while patch tokens are clustered into compact local patterns and matched through optimal transport for fine-grained feature alignment. Balanced optimal transport enforces fixed marginal masses even for local clusters without reliable counterparts, potentially introducing misleading alignment gradients. We therefore introduce confidence-adaptive unbalanced transport to relax these constraints for weakly matched clusters, aiming to reduce unreliable local alignment and improve adversarial transferability. We further study the effect of input transformations and propose visual-invariance augmentation, which applies bidirectional pixel-intensity rescaling and per-channel white-balance adjustment to simulate exposure, contrast, illumination, and color-temperature variations. This strategy encourages adversarial perturbations to generalize across different visual encoders. Extensive experiments on open-source and closed-source MLLMs show that IAU-FOA consistently outperforms state-of-the-art transferable attack methods. Code is available at https://github.com/jiaxiaojunQAQ/IAU-FOA.

Auxiliary-Task Learning (ATL) aims to improve the performance of the target task by leveraging the knowledge obtained from related tasks. Occasionally, learning multiple tasks simultaneously results in lower accuracy than learning only the target task, which is known as negative transfer. This problem is often attributed to the gradient conflicts among tasks, and is frequently tackled by coordinating the task gradients in previous works. However, these optimization-based methods largely overlook the auxiliary-target generalization capability. To better understand the root cause of negative transfer, we experimentally investigate it from both optimization and generalization perspectives. Based on our findings, we introduce ForkMerge, a novel approach that periodically forks the model into multiple branches, automatically searches the varying task weights by minimizing target validation errors, and dynamically merges all branches to filter out detrimental task-parameter updates. On a series of auxiliary-task learning benchmarks, ForkMerge outperforms existing methods and effectively mitigates negative transfer.

Reinforcement Learning with Verifiable Rewards (RLVR) is highly effective for enhancing LLM reasoning, yet recent evidence shows models like Qwen 2.5 achieve significant gains even with spurious or incorrect rewards. We investigate this phenomenon and identify a "Perplexity Paradox": spurious RLVR triggers a divergence where answer-token perplexity drops while prompt-side coherence degrades, suggesting the model is bypassing reasoning in favor of memorization. Using Path Patching, Logit Lens, JSD analysis, and Neural Differential Equations, we uncover a hidden Anchor-Adapter circuit that facilitates this shortcut. We localize a Functional Anchor in the middle layers (L18-20) that triggers the retrieval of memorized solutions, followed by Structural Adapters in later layers (L21+) that transform representations to accommodate the shortcut signal. Finally, we demonstrate that scaling specific MLP keys within this circuit allows for bidirectional causal steering-artificially amplifying or suppressing contamination-driven performance. Our results provide a mechanistic roadmap for identifying and mitigating data contamination in RLVR-tuned models. Code is available at https://github.com/idwts/How-RLVR-Activates-Memorization-Shortcuts.

Aligning foundation models is essential for their safe and trustworthy deployment. However, traditional fine-tuning methods are computationally intensive and require updating billions of model parameters. A promising alternative, alignment via decoding, adjusts the response distribution directly without model updates to maximize a target reward r, thus providing a lightweight and adaptable framework for alignment. However, principled decoding methods rely on oracle access to an optimal Q-function (Q^*), which is often unavailable in practice. Hence, prior SoTA methods either approximate this Q^* using Q^{pi_{sft}} (derived from the reference SFT model) or rely on short-term rewards, resulting in sub-optimal decoding performance. In this work, we propose Transfer Q^*, which implicitly estimates the optimal value function for a target reward r through a baseline model rho_{BL} aligned with a baseline reward rho_{BL} (which can be different from the target reward r). Theoretical analyses of Transfer Q^* provide a rigorous characterization of its optimality, deriving an upper bound on the sub-optimality gap and identifying a hyperparameter to control the deviation from the pre-trained reference SFT model based on user needs. Our approach significantly reduces the sub-optimality gap observed in prior SoTA methods and demonstrates superior empirical performance across key metrics such as coherence, diversity, and quality in extensive tests on several synthetic and real datasets.

Transfer attacks optimize on a surrogate and deploy to a black-box target. While iterative optimization attacks in this paradigm are limited by their per-input cost limits efficiency and scalability due to multistep gradient updates for each input, generative attacks alleviate these by producing adversarial examples in a single forward pass at test time. However, current generative attacks still adhere to optimizing surrogate losses (e.g., feature divergence) and overlook the generator's internal dynamics, underexploring how the generator's internal representations shape transferable perturbations. To address this, we enforce semantic consistency by aligning the early generator's intermediate features to an EMA teacher, stabilizing object-aligned representations and improving black-box transfer without inference-time overhead. To ground the mechanism, we quantify semantic stability as the standard deviation of foreground IoU between cluster-derived activation masks and foreground masks across generator blocks, and observe reduced semantic drift under our method. For more reliable evaluation, we also introduce Accidental Correction Rate (ACR) to separate inadvertent corrections from intended misclassifications, complementing the inherent blind spots in traditional Attack Success Rate (ASR), Fooling Rate (FR), and Accuracy metrics. Across architectures, domains, and tasks, our approach can be seamlessly integrated into existing generative attacks with consistent improvements in black-box transfer, while maintaining test-time efficiency.

Deep neural networks are vulnerable to adversarial examples, dictating the imperativeness to test the model's robustness before deployment. Transfer-based attackers craft adversarial examples against surrogate models and transfer them to victim models deployed in the black-box situation. To enhance the adversarial transferability, structure-based attackers adjust the backpropagation path to avoid the attack from overfitting the surrogate model. However, existing structure-based attackers fail to explore the convolution module in CNNs and modify the backpropagation graph heuristically, leading to limited effectiveness. In this paper, we propose backPropagation pAth Search (PAS), solving the aforementioned two problems. We first propose SkipConv to adjust the backpropagation path of convolution by structural reparameterization. To overcome the drawback of heuristically designed backpropagation paths, we further construct a DAG-based search space, utilize one-step approximation for path evaluation and employ Bayesian Optimization to search for the optimal path. We conduct comprehensive experiments in a wide range of transfer settings, showing that PAS improves the attack success rate by a huge margin for both normally trained and defense models.

While vision-language-action (VLA) models have advanced generalist robotic learning, cross-embodiment transfer remains challenging due to kinematic heterogeneity and the high cost of collecting sufficient real-world demonstrations to support fine-tuning. Existing cross-embodiment policies typically rely on shared-private architectures, which suffer from limited capacity of private parameters and lack explicit adaptation mechanisms. To address these limitations, we introduce MOTIF for efficient few-shot cross-embodiment transfer that decouples embodiment-agnostic spatiotemporal patterns, termed action motifs, from heterogeneous action data. Specifically, MOTIF first learns unified motifs via vector quantization with progress-aware alignment and embodiment adversarial constraints to ensure temporal and cross-embodiment consistency. We then design a lightweight predictor that predicts these motifs from real-time inputs to guide a flow-matching policy, fusing them with robot-specific states to enable action generation on new embodiments. Evaluations across both simulation and real-world environments validate the superiority of MOTIF, which significantly outperforms strong baselines in few-shot transfer scenarios by 6.5% in simulation and 43.7% in real-world settings. Code is available at https://github.com/buduz/MOTIF.

Rapid advancements in large language models (LLMs) have revitalized in LLM-based agents, exhibiting impressive human-like behaviors and cooperative capabilities in various scenarios. However, these agents also bring some exclusive risks, stemming from the complexity of interaction environments and the usability of tools. This paper delves into the safety of LLM-based agents from three perspectives: agent quantity, role definition, and attack level. Specifically, we initially propose to employ a template-based attack strategy on LLM-based agents to find the influence of agent quantity. In addition, to address interaction environment and role specificity issues, we introduce Evil Geniuses (EG), an effective attack method that autonomously generates prompts related to the original role to examine the impact across various role definitions and attack levels. EG leverages Red-Blue exercises, significantly improving the generated prompt aggressiveness and similarity to original roles. Our evaluations on CAMEL, Metagpt and ChatDev based on GPT-3.5 and GPT-4, demonstrate high success rates. Extensive evaluation and discussion reveal that these agents are less robust, prone to more harmful behaviors, and capable of generating stealthier content than LLMs, highlighting significant safety challenges and guiding future research. Our code is available at https://github.com/T1aNS1R/Evil-Geniuses.

Neuro-symbolic systems aim to combine the expressive structure of symbolic logic with the flexibility of neural learning; yet, generative models typically lack mechanisms to enforce declarative constraints at generation time. We propose Logic-Guided Vector Fields (LGVF), a neuro-symbolic framework that injects symbolic knowledge, specified as differentiable relaxations of logical constraints, into flow matching generative models. LGVF couples two complementary mechanisms: (1) a training-time logic loss that penalizes constraint violations along continuous flow trajectories, with weights that emphasize correctness near the target distribution; and (2) an inference-time adjustment that steers sampling using constraint gradients, acting as a lightweight, logic-informed correction to the learned dynamics. We evaluate LGVF on three constrained generation case studies spanning linear, nonlinear, and multi-region feasibility constraints. Across all settings, LGVF reduces constraint violations by 59-82% compared to standard flow matching and achieves the lowest violation rates in each case. In the linear and ring settings, LGVF also improves distributional fidelity as measured by MMD, while in the multi-obstacle setting, we observe a satisfaction-fidelity trade-off, with improved feasibility but increased MMD. Beyond quantitative gains, LGVF yields constraint-aware vector fields exhibiting emergent obstacle-avoidance behavior, routing samples around forbidden regions without explicit path planning.

Action recognition models using deep learning are vulnerable to adversarial examples, which are transferable across other models trained on the same data modality. Existing transferable attack methods face two major challenges: 1) they heavily rely on the assumption that the decision boundaries of the surrogate (a.k.a., source) model and the target model are similar, which limits the adversarial transferability; and 2) their decision boundary difference makes the attack direction uncertain, which may result in the gradient oscillation, weakening the adversarial attack. This motivates us to propose a Background Mixup-induced Temporal Consistency (BMTC) attack method for action recognition. From the input transformation perspective, we design a model-agnostic background adversarial mixup module to reduce the surrogate-target model dependency. In particular, we randomly sample one video from each category and make its background frame, while selecting the background frame with the top attack ability for mixup with the clean frame by reinforcement learning. Moreover, to ensure an explicit attack direction, we leverage the background category as guidance for updating the gradient of adversarial example, and design a temporal gradient consistency loss, which strengthens the stability of the attack direction on subsequent frames. Empirical studies on two video datasets, i.e., UCF101 and Kinetics-400, and one image dataset, i.e., ImageNet, demonstrate that our method significantly boosts the transferability of adversarial examples across several action/image recognition models. Our code is available at https://github.com/mlvccn/BMTC_TransferAttackVid.

We build upon vec2vec, a procedure designed to align text embedding spaces without parallel data. vec2vec finds a near-perfect alignment, but it is expensive and unstable. We present mini-vec2vec, a simple and efficient alternative that requires substantially lower computational cost and is highly robust. Moreover, the learned mapping is a linear transformation. Our method consists of three main stages: a tentative matching of pseudo-parallel embedding vectors, transformation fitting, and iterative refinement. Our linear alternative exceeds the original instantiation of vec2vec by orders of magnitude in efficiency, while matching or exceeding their results. The method's stability and interpretable algorithmic steps facilitate scaling and unlock new opportunities for adoption in new domains and fields.

Vision transformers (ViTs) have been successfully deployed in a variety of computer vision tasks, but they are still vulnerable to adversarial samples. Transfer-based attacks use a local model to generate adversarial samples and directly transfer them to attack a target black-box model. The high efficiency of transfer-based attacks makes it a severe security threat to ViT-based applications. Therefore, it is vital to design effective transfer-based attacks to identify the deficiencies of ViTs beforehand in security-sensitive scenarios. Existing efforts generally focus on regularizing the input gradients to stabilize the updated direction of adversarial samples. However, the variance of the back-propagated gradients in intermediate blocks of ViTs may still be large, which may make the generated adversarial samples focus on some model-specific features and get stuck in poor local optima. To overcome the shortcomings of existing approaches, we propose the Token Gradient Regularization (TGR) method. According to the structural characteristics of ViTs, TGR reduces the variance of the back-propagated gradient in each internal block of ViTs in a token-wise manner and utilizes the regularized gradient to generate adversarial samples. Extensive experiments on attacking both ViTs and CNNs confirm the superiority of our approach. Notably, compared to the state-of-the-art transfer-based attacks, our TGR offers a performance improvement of 8.8% on average.

We introduce Neural Organ Transplantation (NOT), a modular adaptation framework that enables trained transformer layers to function as reusable transferable checkpoints for domain adaptation. Unlike conventional fine-tuning approaches that tightly couple trained parameters to specific model instances and training data, NOT extracts contiguous layer subsets ("donor organs") from pre-trained models, trains them independently on domain-specific data, and saves them as standalone checkpoint files that can be transplanted into compatible recipient models without access to the original training data. Through experiments on three decoder-only transformer architectures spanning 124M to 20B parameters (GPT-2, TinyLlama, and GPT-OSS), we demonstrate that donor transplantation substantially outperforms existing adaptation methods, achieving an order-of-magnitude improvement in perplexity over LoRA while training significantly faster. The method exhibits position dependence, with early insertion positions yielding optimal results. Cross-domain transfer at billion-parameter scale reveals unexpected regularization benefits. These findings demonstrate that transformer middle layers can support efficient modular transfer for decoder-only architectures, enabling privacy-preserving expertise sharing through checkpoint distribution. We note that this approach is currently limited to decoder-only models; preliminary experiments on encoder-based architectures show reduced effectiveness.

Transfer learning is a popular method for tuning pretrained (upstream) models for different downstream tasks using limited data and computational resources. We study how an adversary with control over an upstream model used in transfer learning can conduct property inference attacks on a victim's tuned downstream model. For example, to infer the presence of images of a specific individual in the downstream training set. We demonstrate attacks in which an adversary can manipulate the upstream model to conduct highly effective and specific property inference attacks (AUC score > 0.9), without incurring significant performance loss on the main task. The main idea of the manipulation is to make the upstream model generate activations (intermediate features) with different distributions for samples with and without a target property, thus enabling the adversary to distinguish easily between downstream models trained with and without training examples that have the target property. Our code is available at https://github.com/yulongt23/Transfer-Inference.

Modern large language models (LLMs) are typically secured by auditing data, prompts, and refusal policies, while treating the forward pass as an implementation detail. We show that intermediate activations in decoder-only LLMs form a vulnerable attack surface for behavioral control. Building on recent findings on attention sinks and compression valleys, we identify a high-gain region in the residual stream where small, well-aligned perturbations are causally amplified along the autoregressive trajectory--a Causal Amplification Effect (CAE). We exploit this as an attack surface via Sensitivity-Scaled Steering (SSS), a progressive activation-level attack that combines beginning-of-sequence (BOS) anchoring with sensitivity-based reinforcement to focus a limited perturbation budget on the most vulnerable layers and tokens. We show that across multiple open-weight models and four behavioral axes, SSS induces large shifts in evil, hallucination, sycophancy, and sentiment while preserving high coherence and general capabilities, turning activation steering into a concrete security concern for white-box and supply-chain LLM deployments.

Vision-Language Model (VLM) based Web Agents represent a significant step towards automating complex tasks by simulating human-like interaction with websites. However, their deployment in uncontrolled web environments introduces significant security vulnerabilities. Existing research on adversarial environmental injection attacks often relies on unrealistic assumptions, such as direct HTML manipulation, knowledge of user intent, or access to agent model parameters, limiting their practical applicability. In this paper, we propose AdInject, a novel and real-world black-box attack method that leverages the internet advertising delivery to inject malicious content into the Web Agent's environment. AdInject operates under a significantly more realistic threat model than prior work, assuming a black-box agent, static malicious content constraints, and no specific knowledge of user intent. AdInject includes strategies for designing malicious ad content aimed at misleading agents into clicking, and a VLM-based ad content optimization technique that infers potential user intents from the target website's context and integrates these intents into the ad content to make it appear more relevant or critical to the agent's task, thus enhancing attack effectiveness. Experimental evaluations demonstrate the effectiveness of AdInject, attack success rates exceeding 60% in most scenarios and approaching 100% in certain cases. This strongly demonstrates that prevalent advertising delivery constitutes a potent and real-world vector for environment injection attacks against Web Agents. This work highlights a critical vulnerability in Web Agent security arising from real-world environment manipulation channels, underscoring the urgent need for developing robust defense mechanisms against such threats. Our code is available at https://github.com/NicerWang/AdInject.

Subliminal learning refers to a student language model acquiring a teacher's traits (e.g. a system-prompted preference for owls) when fine-tuned on the teacher's outputs, despite the outputs being semantically unrelated to those traits. It remains poorly understood how data without semantic meaning can transfer specific semantic traits. In this work, we show that subliminal learning is mediated by a single steering vector, i.e. a vector added to the model's activations. Across two open-source models, we find that the teacher's system prompt is well approximated by a steering vector, and that the student's behavior is driven by learning an aligned vector over fine-tuning. System prompts that are not well approximated by steering vectors are not subliminally learned. This is a special case of steering vector distillation, in which a student trained on the outputs of a steered teacher learns to imitate that steering. We demonstrate steering vector distillation on a range of semantic and random vectors. Adding a semantic vector to a model's activations can have both model-independent and model-specific (i.e. non-semantic) effects on its behavior, so generated data that is non-semantic can transmit a vector with semantic effects, enabling subliminal learning. This also explains why subliminal learning does not transfer between models. We find that adaptive optimizers are necessary for subliminal learning in language models: activation gradients on steered data carry a small but consistent component along the steering direction, and non-adaptive optimizers impede this by allowing outlier gradients to dominate.

We show that when large language models learn to reward hack on production RL environments, this can result in egregious emergent misalignment. We start with a pretrained model, impart knowledge of reward hacking strategies via synthetic document finetuning or prompting, and train on a selection of real Anthropic production coding environments. Unsurprisingly, the model learns to reward hack. Surprisingly, the model generalizes to alignment faking, cooperation with malicious actors, reasoning about malicious goals, and attempting sabotage when used with Claude Code, including in the codebase for this paper. Applying RLHF safety training using standard chat-like prompts results in aligned behavior on chat-like evaluations, but misalignment persists on agentic tasks. Three mitigations are effective: (i) preventing the model from reward hacking; (ii) increasing the diversity of RLHF safety training; and (iii) "inoculation prompting", wherein framing reward hacking as acceptable behavior during training removes misaligned generalization even when reward hacking is learned.

Understanding social interactions from visual cues is a fundamental challenge for a socially competent AI. While powerful pre-trained vision-language models (VLMs) have shown remarkable general capabilities, they surprisingly struggle to unify and learn multiple social perception tasks simultaneously, often exhibiting negative transfer. We identify that this negative transfer stems from a critical issue we term "social degradation," whereby the general visual-linguistic pre-training process of VLMs impairs the visual encoder's ability to represent nuanced social information. We investigate this behavior further under two lenses: decodability through linear representation probing and compatibility through gradient conflict analysis, revealing that both play a role in the degradation, especially the former, which is significantly compromised in the VLM pre-training process. To address these issues, we propose SocialFusion, a unified framework that learns a minimal connection between a frozen visual encoder and a language model. Compared with existing VLMs, it exhibits positive transfer across all five social tasks, leveraging synergies between them to enhance overall performance and achieves comparable performance to task-specific state-of-the-art models on various benchmarks. Our findings suggest that current VLM pre-training strategies may be detrimental to acquiring general social competence and highlight the need for more socially-aware training paradigms.

Jailbreaking attacks can effectively manipulate open-source large language models (LLMs) to produce harmful responses. However, these attacks exhibit limited transferability, failing to disrupt proprietary LLMs consistently. To reliably identify vulnerabilities in proprietary LLMs, this work investigates the transferability of jailbreaking attacks by analysing their impact on the model's intent perception. By incorporating adversarial sequences, these attacks can redirect the source LLM's focus away from malicious-intent tokens in the original input, thereby obstructing the model's intent recognition and eliciting harmful responses. Nevertheless, these adversarial sequences fail to mislead the target LLM's intent perception, allowing the target LLM to refocus on malicious-intent tokens and abstain from responding. Our analysis further reveals the inherent distributional dependency within the generated adversarial sequences, whose effectiveness stems from overfitting the source LLM's parameters, resulting in limited transferability to target LLMs. To this end, we propose the Perceived-importance Flatten (PiF) method, which uniformly disperses the model's focus across neutral-intent tokens in the original input, thus obscuring malicious-intent tokens without relying on overfitted adversarial sequences. Extensive experiments demonstrate that PiF provides an effective and efficient red-teaming evaluation for proprietary LLMs.

Transfer learning leverages feature representations of deep neural networks (DNNs) pretrained on source tasks with rich data to empower effective finetuning on downstream tasks. However, the pretrained models are often prohibitively large for delivering generalizable representations, which limits their deployment on edge devices with constrained resources. To close this gap, we propose a new transfer learning pipeline, which leverages our finding that robust tickets can transfer better, i.e., subnetworks drawn with properly induced adversarial robustness can win better transferability over vanilla lottery ticket subnetworks. Extensive experiments and ablation studies validate that our proposed transfer learning pipeline can achieve enhanced accuracy-sparsity trade-offs across both diverse downstream tasks and sparsity patterns, further enriching the lottery ticket hypothesis.

As Large Language Models (LLMs) grow increasingly powerful, multi-agent systems are becoming more prevalent in modern AI applications. Most safety research, however, has focused on vulnerabilities in single-agent LLMs. These include prompt injection attacks, where malicious prompts embedded in external content trick the LLM into executing unintended or harmful actions, compromising the victim's application. In this paper, we reveal a more dangerous vector: LLM-to-LLM prompt injection within multi-agent systems. We introduce Prompt Infection, a novel attack where malicious prompts self-replicate across interconnected agents, behaving much like a computer virus. This attack poses severe threats, including data theft, scams, misinformation, and system-wide disruption, all while propagating silently through the system. Our extensive experiments demonstrate that multi-agent systems are highly susceptible, even when agents do not publicly share all communications. To address this, we propose LLM Tagging, a defense mechanism that, when combined with existing safeguards, significantly mitigates infection spread. This work underscores the urgent need for advanced security measures as multi-agent LLM systems become more widely adopted.

Deepfake detectors remain vulnerable to transfer-based black-box attacks, in which adversarial examples are generated on a source surrogate model and transferred to a target model, unknown to the attacker. Yet how source--target compatibility shapes attack success remains poorly understood. Prior studies evaluate limited detector pools and rarely disentangle architectural from training factors. We conduct a controlled evaluation of adversarial transferability across 60 detectors spanning six backbones, two pretraining regimes, and five training-data configurations, using two attack procedures: AutoAttack (AA) and the Carlini--Wagner attack with Expectation over Transformation (CW--EOT). Matched comparisons reveal significantly higher transfer when source and target share an exact backbone, architecture family, pretraining regime, or training data. This compatibility structure is attack-dependent: exact backbone compatibility has the largest effect under AA, whereas shared pretraining and training data have the largest effects under CW--EOT. When transfer is averaged across non-target sources, mean attack success rate (ASR) is 7.21% under AA and 19.52% under CW--EOT. By contrast, a multi-source oracle combining both attacks attains a \(64.48\%\) mean ASR after excluding exact backbone and training-data matches, showing that source averaging can substantially understate target vulnerability. We release 240,000 adversarially perturbed images, complete pairwise transfer results, detector configurations, and evaluation code. These findings establish source--target compatibility and source-model selection as central dimensions of credible transfer-based black-box robustness evaluation.

Recently, text-guided scalable vector graphics (SVGs) synthesis has shown promise in domains such as iconography and sketch. However, existing text-to-SVG generation methods lack editability and struggle with visual quality and result diversity. To address these limitations, we propose a novel text-guided vector graphics synthesis method called SVGDreamer. SVGDreamer incorporates a semantic-driven image vectorization (SIVE) process that enables the decomposition of synthesis into foreground objects and background, thereby enhancing editability. Specifically, the SIVE process introduce attention-based primitive control and an attention-mask loss function for effective control and manipulation of individual elements. Additionally, we propose a Vectorized Particle-based Score Distillation (VPSD) approach to tackle the challenges of color over-saturation, vector primitives over-smoothing, and limited result diversity in existing text-to-SVG generation methods. Furthermore, on the basis of VPSD, we introduce Reward Feedback Learning (ReFL) to accelerate VPSD convergence and improve aesthetic appeal. Extensive experiments have been conducted to validate the effectiveness of SVGDreamer, demonstrating its superiority over baseline methods in terms of editability, visual quality, and diversity.

Recent advances in operating system (OS) agents enable vision-language models to interact directly with the graphical user interface of an OS. These multimodal OS agents autonomously perform computer-based tasks in response to a single prompt via application programming interfaces (APIs). Such APIs typically support low-level operations, including mouse clicks, keyboard inputs, and screenshot captures. We introduce a novel attack vector: malicious image patches (MIPs) that have been adversarially perturbed so that, when captured in a screenshot, they cause an OS agent to perform harmful actions by exploiting specific APIs. For instance, MIPs embedded in desktop backgrounds or shared on social media can redirect an agent to a malicious website, enabling further exploitation. These MIPs generalise across different user requests and screen layouts, and remain effective for multiple OS agents. The existence of such attacks highlights critical security vulnerabilities in OS agents, which should be carefully addressed before their widespread adoption.

Indirect prompt injection poses a critical threat to LLM agents that interact with untrusted external data, yet automated attack methods--proven effective for jailbreaking--remain underexplored in realistic agentic settings. We present a comprehensive empirical evaluation of automated prompt injection attacks against LLM agents, adapting both white-box (GCG) and black-box (TAP) methods to the agentic setting within the AgentDojo framework. We evaluate across 80 task pairs spanning four domains and multiple models, and find that black-box optimization substantially outperforms gradient-based methods, a gap we attribute to GCG's optimization instability under reasonable compute budgets. We also find that TAP's effectiveness depends on the attacker model, as both general capability and safety tuning affect attack success--stronger models produce more effective injections, while safety-tuned attackers can refuse to generate adversarial prompts. Task-universal attacks transfer effectively to unseen tasks and out-of-distribution domains, but attacks optimized on smaller open-source models do not transfer to frontier models like GPT-5. These findings highlight automated prompt injection as a credible but model-dependent threat, with significant barriers remaining for model-agnostic exploitation.

Large Language Models (LLMs), such as OpenAI's o1-series have demonstrated compelling capabilities for complex reasoning tasks via the extended chain-of-thought (CoT) reasoning mechanism. However, recent studies reveal substantial redundancy in the CoT reasoning traces, which not only increases inference latency but also negatively impacts model performance by diverting attention to unnecessary reasoning paths. To address this issue, we investigate the internal reasoning structures of LLMs and categorize them into three primary thought types: execution, reflection, and transition thoughts. Moreover, our analysis reveals that excessive reflection and transition thoughts are strongly correlated with failure cases and these thought categories exhibit clear separation in the latent space. Based on these, we introduce SEAL (Steerable reasoning calibration), a training-free approach that seamlessly calibrates the CoT process, improving accuracy while demonstrating significant efficiency gains. SEAL consists of an offline stage for extracting the reasoning steering vector in the latent space, followed by an on-the-fly calibration of the reasoning trace through representation intervention using the steering vector. Notably, the steering vector exhibits strong transferability across various tasks. Extensive experiments across multiple models (DeepSeek-R1-Distill and QwQ-32B-Preview) and benchmarks (Math500, GSM8K, LiveCodeBench) validate the effectiveness of SEAL, up to a 11% improvement in accuracy while reducing reasoning tokens by 11.8% to 50.4%. Our code is publicly available at https://github.com/VITA-Group/SEAL.

An established way to improve the transferability of black-box evasion attacks is to craft the adversarial examples on an ensemble-based surrogate to increase diversity. We argue that transferability is fundamentally related to uncertainty. Based on a state-of-the-art Bayesian Deep Learning technique, we propose a new method to efficiently build a surrogate by sampling approximately from the posterior distribution of neural network weights, which represents the belief about the value of each parameter. Our extensive experiments on ImageNet, CIFAR-10 and MNIST show that our approach improves the success rates of four state-of-the-art attacks significantly (up to 83.2 percentage points), in both intra-architecture and inter-architecture transferability. On ImageNet, our approach can reach 94% of success rate while reducing training computations from 11.6 to 2.4 exaflops, compared to an ensemble of independently trained DNNs. Our vanilla surrogate achieves 87.5% of the time higher transferability than three test-time techniques designed for this purpose. Our work demonstrates that the way to train a surrogate has been overlooked, although it is an important element of transfer-based attacks. We are, therefore, the first to review the effectiveness of several training methods in increasing transferability. We provide new directions to better understand the transferability phenomenon and offer a simple but strong baseline for future work.

As the application space of language models continues to evolve, a natural question to ask is how we can quickly adapt models to new tasks. We approach this classic question from a continual learning perspective, in which we aim to continue fine-tuning models trained on past tasks on new tasks, with the goal of "transferring" relevant knowledge. However, this strategy also runs the risk of doing more harm than good, i.e., negative transfer. In this paper, we construct a new benchmark of task sequences that target different possible transfer scenarios one might face, such as a sequence of tasks with high potential of positive transfer, high potential for negative transfer, no expected effect, or a mixture of each. An ideal learner should be able to maximally exploit information from all tasks that have any potential for positive transfer, while also avoiding the negative effects of any distracting tasks that may confuse it. We then propose a simple, yet effective, learner that satisfies many of our desiderata simply by leveraging a selective strategy for initializing new models from past task checkpoints. Still, limitations remain, and we hope this benchmark can help the community to further build and analyze such learners.

Recently, language models like Llama 3.1 Instruct have become increasingly capable of agentic behavior, enabling them to perform tasks requiring short-term planning and tool use. In this study, we apply refusal-vector ablation to Llama 3.1 70B and implement a simple agent scaffolding to create an unrestricted agent. Our findings imply that these refusal-vector ablated models can successfully complete harmful tasks, such as bribing officials or crafting phishing attacks, revealing significant vulnerabilities in current safety mechanisms. To further explore this, we introduce a small Safe Agent Benchmark, designed to test both harmful and benign tasks in agentic scenarios. Our results imply that safety fine-tuning in chat models does not generalize well to agentic behavior, as we find that Llama 3.1 Instruct models are willing to perform most harmful tasks without modifications. At the same time, these models will refuse to give advice on how to perform the same tasks when asked for a chat completion. This highlights the growing risk of misuse as models become more capable, underscoring the need for improved safety frameworks for language model agents.

We show that robustness to post-training quantization (PTQ) is a transferable direction in weight space. We call this direction the quantization vector: extracted from a donor task by simple weight-space arithmetic, it can be used to patch a receiver model and improve post-PTQ Top-1 accuracy by up to 60 points in a 3-bit setting, without receiver-side quantization-aware training (QAT). Because the method requires no receiver training data, it provides a zero-shot, low-cost alternative to QAT for extremely low-bit deployment. Across four ViT scales and 22 image classification tasks, donor quantization vectors often yield substantial gains even when donor and receiver tasks differ markedly. We further prove rigorously that quantization vectors are well-defined and do not suffer from reparameterization symmetries, and provide a local geometric account of their effect. Together, these results suggest that quantization robustness can be partially isolated, reused, and transferred through simple weight-space algebra.

Jailbreaking attacks can effectively induce unsafe behaviors in Large Language Models (LLMs); however, the transferability of these attacks across different models remains limited. This study aims to understand and enhance the transferability of gradient-based jailbreaking methods, which are among the standard approaches for attacking white-box models. Through a detailed analysis of the optimization process, we introduce a novel conceptual framework to elucidate transferability and identify superfluous constraints-specifically, the response pattern constraint and the token tail constraint-as significant barriers to improved transferability. Removing these unnecessary constraints substantially enhances the transferability and controllability of gradient-based attacks. Evaluated on Llama-3-8B-Instruct as the source model, our method increases the overall Transfer Attack Success Rate (T-ASR) across a set of target models with varying safety levels from 18.4% to 50.3%, while also improving the stability and controllability of jailbreak behaviors on both source and target models.

We propose transferability from Large Geometric Vicinity (LGV), a new technique to increase the transferability of black-box adversarial attacks. LGV starts from a pretrained surrogate model and collects multiple weight sets from a few additional training epochs with a constant and high learning rate. LGV exploits two geometric properties that we relate to transferability. First, models that belong to a wider weight optimum are better surrogates. Second, we identify a subspace able to generate an effective surrogate ensemble among this wider optimum. Through extensive experiments, we show that LGV alone outperforms all (combinations of) four established test-time transformations by 1.8 to 59.9 percentage points. Our findings shed new light on the importance of the geometry of the weight space to explain the transferability of adversarial examples.

Recently, text-guided scalable vector graphics (SVG) synthesis has demonstrated significant potential in domains such as iconography and sketching. However, SVGs generated from existing Text-to-SVG methods often lack editability and exhibit deficiencies in visual quality and diversity. In this paper, we propose a novel text-guided vector graphics synthesis method to address these limitations. To enhance the editability of output SVGs, we introduce a Hierarchical Image VEctorization (HIVE) framework that operates at the semantic object level and supervises the optimization of components within the vector object. This approach facilitates the decoupling of vector graphics into distinct objects and component levels. Our proposed HIVE algorithm, informed by image segmentation priors, not only ensures a more precise representation of vector graphics but also enables fine-grained editing capabilities within vector objects. To improve the diversity of output SVGs, we present a Vectorized Particle-based Score Distillation (VPSD) approach. VPSD addresses over-saturation issues in existing methods and enhances sample diversity. A pre-trained reward model is incorporated to re-weight vector particles, improving aesthetic appeal and enabling faster convergence. Additionally, we design a novel adaptive vector primitives control strategy, which allows for the dynamic adjustment of the number of primitives, thereby enhancing the presentation of graphic details. Extensive experiments validate the effectiveness of the proposed method, demonstrating its superiority over baseline methods in terms of editability, visual quality, and diversity. We also show that our new method supports up to six distinct vector styles, capable of generating high-quality vector assets suitable for stylized vector design and poster design. Code and demo will be released at: http://ximinng.github.io/SVGDreamerV2Project/

Scaling laws for large language models depend critically on the optimizer and parameterization. Existing hyperparameter transfer laws are mainly developed for first-order optimizers, and they do not structurally prevent training instability at scale. Recent hypersphere optimization methods constrain weight matrices to a fixed-norm hypersphere, offering a promising alternative for more stable scaling. We introduce HyperP (Hypersphere Parameterization), the first framework for transferring optimal learning rates across model width, depth, training tokens, and Mixture-of-Experts (MoE) granularity under the Frobenius-sphere constraint with the Muon optimizer. We prove that weight decay is a first-order no-op on the Frobenius sphere, show that Depth-μP remains necessary, and find that the optimal learning rate follows the same data-scaling power law with the "magic exponent" 0.32 previously observed for AdamW. A single base learning rate tuned at the smallest scale transfers across all compute budgets under HyperP, yielding 1.58times compute efficiency over a strong Muon baseline at 6times10^{21} FLOPs. Moreover, HyperP delivers transferable stability: all monitored instability indicators, including Z-values, output RMS, and activation outliers, remain bounded and non-increasing under training FLOPs scaling. We also propose SqrtGate, an MoE gating mechanism derived from the hypersphere constraint that preserves output RMS across MoE granularities for improved granularity scaling, and show that hypersphere optimization enables substantially larger auxiliary load-balancing weights, yielding both strong performance and good expert balance. We release our training codebase at https://github.com/microsoft/ArchScale.

Modern retrieval-augmented generation (RAG) systems convert sensitive content into high-dimensional embeddings and store them in vector databases that treat the resulting numerical artifacts as opaque. Major vector-store products do not provide native controls for embedding integrity, ingestion-time distributional anomaly detection, or cryptographic provenance attestation. We show this opens a class of steganographic exfiltration attacks: an attacker with write access to the ingestion pipeline can hide payload data inside embeddings using simple post-embedding perturbations (noise injection, rotation, scaling, offset, fragmentation, and combinations thereof) while preserving the surface-level retrieval behavior the RAG system exposes to legitimate users. We evaluate these techniques across a synthetic-PII corpus on text-embedding-3-large, four locally hosted open embedding models, a cross-corpus replication on BEIR NFCorpus and a Quora subset (over 26,000 chunks combined), seven vector-store configurations, an adaptive-attacker variant of the detector evaluation, and a paraphrased-query retrieval benchmark. Distribution-shifting perturbations are often caught by simple anomaly detectors; small-angle orthogonal rotation defeats distribution-based detection across every (model, corpus) pair tested. A disjoint-Givens rotation encoder gives a closed-form per-vector capacity ceiling of floor(d/2) * b bits, but real embedding manifolds impose a capacity-detectability trade-off, and the retrieval-preserving operating point sits well below it. We propose VectorPin, a cryptographic provenance protocol that pins each embedding to its source content and producing model via an Ed25519 signature over a canonical byte representation. Any post-embedding modification breaks signature verification. Embedding-level integrity is a deployable, standardizable control that closes this attack class.

Empirical scaling laws prescribe how to allocate parameters, data, and compute, while maximal-update parameterization (muP) enables learning-rate transfer across widths by equalizing early-time update magnitudes. However, in modern scale-invariant architectures, training quickly enters an optimizer-governed steady state where normalization layers create backward scale sensitivity and the effective learning rate becomes width dependent, degrading muP transfer. We address this by introducing a weight-decay scaling rule for AdamW that preserves sublayer gain across widths. Empirically, the singular-value spectrum of each matrix parameter scales in norm as eta/lambda with an approximately invariant shape; under width scaling d, we observe that the top singular value scales approximately as eta/lambdacdot d^{0.75}. Combining this observation with the muP learning-rate rule eta_2propto d^{-1} for matrix-like parameters implies an empirical weight-decay scaling rule lambda_2propto d that approximately keeps sublayer gains width invariant. Together with vector-like parameters trained at eta_1=Theta_d(1) and lambda_1=0, this yields zero-shot transfer of both learning rate and weight decay from proxy to target widths, removing per-width sweeps. We validate the rule on LLaMA-style Transformers and in a minimal synthetic setting, and we provide a simple diagnostic, matching top singular values, to check sublayer-gain invariance. Our results extend muP beyond the near-init regime by explicitly controlling steady-state scales set by the optimizer, offering a practical recipe for width-robust hyperparameter transfer under AdamW.

Understanding how aligned LLMs internally represent safety is critical for diagnosing alignment vulnerabilities, as it explains why jailbreaks succeed and informs the design of robust alignment strategies. Prior work shows that aligned LLMs encode harmfulness and refusal as separable directions in the residual stream at prompt-side token positions. We show that jailbreaks succeed at prompt encoding by suppressing either the refusal or harmfulness direction before any token is generated, with distinct attack classes occupying separable regions of the harmfulness-refusal plane. Extending the analysis to response-token positions, we find that the model recognizes harmful content while it is generating that content, even when it failed to recognize the input as harmful at the prompt side. Motivated by our findings, we introduce HARC (Harmfulness-And-Refusal Coupling), a fine-tuning method that pairs the two directions across both prompt and response positions. Since the intervention is confined to the harmfulness-refusal subspace, it leaves the rest of the residual stream intact and does not degrade general capability or inflate over-refusal. Across extensive experiments, HARC achieves the strongest robustness-capability-usability trade-off among six baselines spanning the major training-time and inference-time safety methods. The harmfulness and refusal directions at prompt and response positions transfer across the five model families and two scales we tested without architecture-specific tuning.

AI evaluation can support the wrong inference when an in-domain benchmark success does not survive distribution shift, or when the benchmark endpoint is entangled with a design factor. We study this problem in CRISPRi perturbation-effect prediction, evaluating a frozen Geneformer representation under a locked, pre-registered protocol: heads and model selection were frozen before test evaluation; the protocol required external outcome labels to remain withheld until final unblinding; and analysis-governing decisions were fixed before the evaluations they govern. In-distribution on the Virtual Cell Challenge (VCC), the frozen representation carries measurable predictive information beyond a dimension-matched random-feature control (Delta R^2 = +0.1645, 95% CI [+0.1375, +0.1920]), satisfying the pre-registered informativeness gate required before interpreting transfer. It then fails zero-shot transfer on both external screens (Spearman rho = -0.139 and -0.267), lying below that control on each. Adding a predefined magnitude block improves the representation externally (Delta rho = +0.032 and +0.143) but, under the frozen primary head, does not rescue transfer: both remain negative. A pre-registered, count-adjusted max-response secondary is positively associated with the outcome on both screens; we report it as correlational and secondary, not as a recovered magnitude signal. Finally, the VCC endpoint is strongly sample-size associated: a count-only linear model reaches R^2 = +0.4325, versus +0.2589 for the four magnitude scalars; adding those scalars to cell count improves R^2 by only +0.0017, so much of the aggregate-magnitude signal overlaps with cell count. This case study shows how locking the evaluation, harmonizing the measured endpoint, and separating primary from secondary evidence can change the inference supported by an AI benchmark.

Safety alignment instills in Large Language Models (LLMs) a critical capacity to refuse malicious requests. Prior works have modeled this refusal mechanism as a single linear direction in the activation space. We posit that this is an oversimplification that conflates two functionally distinct neural processes: the detection of harm and the execution of a refusal. In this work, we deconstruct this single representation into a Harm Detection Direction and a Refusal Execution Direction. Leveraging this fine-grained model, we introduce Differentiated Bi-Directional Intervention (DBDI), a new white-box framework that precisely neutralizes the safety alignment at critical layer. DBDI applies adaptive projection nullification to the refusal execution direction while suppressing the harm detection direction via direct steering. Extensive experiments demonstrate that DBDI outperforms prominent jailbreaking methods, achieving up to a 97.88\% attack success rate on models such as Llama-2. By providing a more granular and mechanistic framework, our work offers a new direction for the in-depth understanding of LLM safety alignment.

While open-weight models have enjoyed steady progress in capabilities and wide adoption across multiple domains, their safety remains an important concern. One key feature is the ability to refuse or deflect harmful, malicious, or insensitive prompts. In this paper, we expose safety vulnerabilities across six common open-weight LLMs of various sizes that consistently lead to harmful or unsafe responses on the JailbreakBench benchmark dataset. Our proposed attack, Perturbed Embedding Vector (PEV), is a simple and fast "jailbreaking" technique that is cheaper than prior approaches, which typically require gradient computations, per-prompt optimizations, or altering internal weights of the models. PEV just adds independent Gaussian noise in the embedding vector representations of the prompt, with no need for further manipulations. To generate unsafe responses, we repeatedly sample additive noise from this distribution. In our experiments, we observe that the average compute cost to get the first successful attack is up to an order of magnitude less than previous attacks. The first successful jailbreak on a new prompt typically arrives within one minute on every tested model, and PEV generates unsafe responses across all models for all prompts in JailbreakBench. No other tested method achieves such results, despite them taking longer to run. More broadly, we believe that understanding the behavior of LLMs under perturbations in the embedding vectors is an important research direction: while perturbations constitute a major security risk, they can also serve as a valuable tool for exploring the dynamical behavior of such models.

Ensemble-based attacks have been proven to be effective in enhancing adversarial transferability by aggregating the outputs of models with various architectures. However, existing research primarily focuses on refining ensemble weights or optimizing the ensemble path, overlooking the exploration of ensemble models to enhance the transferability of adversarial attacks. To address this gap, we propose applying adversarial augmentation to the surrogate models, aiming to boost overall generalization of ensemble models and reduce the risk of adversarial overfitting. Meanwhile, observing that ensemble Vision Transformers (ViTs) gain less attention, we propose ViT-EnsembleAttack based on the idea of model adversarial augmentation, the first ensemble-based attack method tailored for ViTs to the best of our knowledge. Our approach generates augmented models for each surrogate ViT using three strategies: Multi-head dropping, Attention score scaling, and MLP feature mixing, with the associated parameters optimized by Bayesian optimization. These adversarially augmented models are ensembled to generate adversarial examples. Furthermore, we introduce Automatic Reweighting and Step Size Enlargement modules to boost transferability. Extensive experiments demonstrate that ViT-EnsembleAttack significantly enhances the adversarial transferability of ensemble-based attacks on ViTs, outperforming existing methods by a substantial margin. Code is available at https://github.com/Trustworthy-AI-Group/TransferAttack.

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.

Scalable Vector Graphics (SVG) offer a powerful format for representing visual designs as interpretable code. Recent advances in vision-language models (VLMs) have enabled high-quality SVG generation by framing the problem as a code generation task and leveraging large-scale pretraining. VLMs are particularly suitable for this task as they capture both global semantics and fine-grained visual patterns, while transferring knowledge across vision, natural language, and code domains. However, existing VLM approaches often struggle to produce faithful and efficient SVGs because they never observe the rendered images during training. Although differentiable rendering for autoregressive SVG code generation remains unavailable, rendered outputs can still be compared to original inputs, enabling evaluative feedback suitable for reinforcement learning (RL). We introduce RLRF(Reinforcement Learning from Rendering Feedback), an RL method that enhances SVG generation in autoregressive VLMs by leveraging feedback from rendered SVG outputs. Given an input image, the model generates SVG roll-outs that are rendered and compared to the original image to compute a reward. This visual fidelity feedback guides the model toward producing more accurate, efficient, and semantically coherent SVGs. RLRF significantly outperforms supervised fine-tuning, addressing common failure modes and enabling precise, high-quality SVG generation with strong structural understanding and generalization.

Recent advancements in AI safety have led to increased efforts in training and red-teaming large language models (LLMs) to mitigate unsafe content generation. However, these safety mechanisms may not be comprehensive, leaving potential vulnerabilities unexplored. This paper introduces MathPrompt, a novel jailbreaking technique that exploits LLMs' advanced capabilities in symbolic mathematics to bypass their safety mechanisms. By encoding harmful natural language prompts into mathematical problems, we demonstrate a critical vulnerability in current AI safety measures. Our experiments across 13 state-of-the-art LLMs reveal an average attack success rate of 73.6\%, highlighting the inability of existing safety training mechanisms to generalize to mathematically encoded inputs. Analysis of embedding vectors shows a substantial semantic shift between original and encoded prompts, helping explain the attack's success. This work emphasizes the importance of a holistic approach to AI safety, calling for expanded red-teaming efforts to develop robust safeguards across all potential input types and their associated risks.

In this paper, we present techniques used to implement our portable vectorized library of C standard mathematical functions written entirely in C language. In order to make the library portable while maintaining good performance, intrinsic functions of vector extensions are abstracted by inline functions or preprocessor macros. We implemented the functions so that they can use sub-features of vector extensions such as fused multiply-add, mask registers and extraction of mantissa. In order to make computation with SIMD instructions efficient, the library only uses a small number of conditional branches, and all the computation paths are vectorized. We devised a variation of the Payne-Hanek argument reduction for trigonometric functions and a floating point remainder, both of which are suitable for vector computation. We compare the performance of our library to Intel SVML.

As large language models (LLMs) are increasingly deployed in alignment-sensitive contexts, activation steering has emerged as a lightweight, inference-time alternative to fine-tuning methods (e.g., RLHF, DPO) for behavioral control. However, existing work typically validates steering on isolated behaviors, leaving it unclear whether steering vectors encode coherent semantic structure or merely exploit behavior-specific shortcuts. We investigate whether the latent geometry of LLM steering vectors reflects theory-specified structure in human values and morality. Using Schwartz's Theory of Basic Human Values as our primary fine-grained framework, we introduce a 26K-sample benchmark covering 20 human values and analyze distribution-driven methods (e.g., CAA, SphericalSteer, ODESteer) and behavior-centric approaches (e.g., COLD-Steer, BiPO) across diverse model families and sizes. We find that distribution-driven methods recover human value topologies aligned with theoretical predictions (Spearman ρ up to 0.51, p < 10^{-13}). In contrast, behavior-centric methods achieve comparable steering performance but show little correlation with the expected value geometry. Geometric fidelity improves with model scale but drops after instruction tuning. Finally, better geometric alignment also leads to more human-consistent transfer across values: steering one value correctly lifts compatible values and suppresses opposing ones. Code and data are available at: https://github.com/DeepRCL/Steering_Geometry.

We identify Harmfulness Propagation Dynamics (HPD): for harmful prompts, the projection of the last-token hidden state onto a learned harm direction rises monotonically with transformer depth, whereas benign prompts remain flat or oscillatory. This cross-layer signature reflects harmful intent as a progressively resolved semantic property: surface form appears early, while pragmatic intent consolidates later, making the trajectory shape more informative than any single-layer snapshot. Moreover, LDA-based harm directions, learned per layer, remain stable across random splits (pairwise cosine similarity >0.97), supporting the projection sequence as a reproducible structured signal. Building on HPD, we introduce \herald{} (Harmful Encoding Recognition via Activation Layer Dynamics). This lightweight input moderator extracts a seven-dimensional feature record, slope, curvature, monotonicity, onset layer, and related statistics from the cross-layer projection sequence and classifies it with a 288-parameter MLP. stores one d-dimensional direction per layer (262\,KB for a 32-layer, d{=}4096 model), requires no gradient computation during training, and adds only 2.6{times}10^{-6} prefill FLOPs at inference. Across eight prompt-harmfulness benchmarks and four model families, achieves an average F1 of 89.3 on OLMo2-7B, surpassing all tested guard models on adversarial jailbreak detection (98.4 vs.\ 96.9 F1) and outperforming prior latent-based methods by 2.3-4.1 F1 points on every backbone. Per-instance trajectories provide machine-readable audit records that reveal when and how harmfulness emerges, offering an interpretability advantage over single-layer approaches.

Emergent misalignment (EM) is the surprising tendency of language models to become broadly misaligned after fine-tuning on narrowly misaligned examples. While EM has been extensively studied in the supervised fine-tuning (SFT) setting, evidence that it also arises from reinforcement learning (RL) is limited to large, closed-source models, leaving the phenomenon expensive to study and difficult to reproduce. We characterize EM from RL in small, off-the-shelf open-weight models along three axes. First, we show that rewarding narrow, overtly misaligned behavior produces substantially higher general-domain misalignment than sample-matched SFT. Second, we show that EM from RL can be induced by reward signals that could plausibly arise naturally, such as unpopular aesthetic preferences or poor rhetorical appeals. Third, we evaluate in-training mitigations developed for SFT-induced EM and find that they broadly transfer, with preventive steering with persona vectors, interleaving safety data and inoculation prompting all performing well.

The Identifiable Victim Effect (IVE) - the tendency to allocate greater resources to a specific, narratively described victim than to a statistically characterized group facing equivalent hardship - is one of the most robust findings in moral psychology and behavioural economics. As large language models (LLMs) assume consequential roles in humanitarian triage, automated grant evaluation, and content moderation, a critical question arises: do these systems inherit the affective irrationalities present in human moral reasoning? We present the first systematic, large-scale empirical investigation of the IVE in LLMs, comprising N=51,955 validated API trials across 16 frontier models spanning nine organizational lineages (Google, Anthropic, OpenAI, Meta, DeepSeek, xAI, Alibaba, IBM, and Moonshot). Using a suite of ten experiments - porting and extending canonical paradigms from Small et al. (2007) and Kogut and Ritov (2005) - we find that the IVE is prevalent but strongly modulated by alignment training. Instruction-tuned models exhibit extreme IVE (Cohen's d up to 1.56), while reasoning-specialized models invert the effect (down to d=-0.85). The pooled effect (d=0.223, p=2e-6) is approximately twice the single-victim human meta-analytic baseline (dapprox0.10) reported by Lee and Feeley (2016) - and likely exceeds the overall human pooled effect by a larger margin, given that the group-victim human effect is near zero. Standard Chain-of-Thought (CoT) prompting - contrary to its role as a deliberative corrective - nearly triples the IVE effect size (from d=0.15 to d=0.41), while only utilitarian CoT reliably eliminates it. We further document psychophysical numbing, perfect quantity neglect, and marginal in-group/out-group cultural bias, with implications for AI deployment in humanitarian and ethical decision-making contexts.

Vision-language models (VLMs) have improved significantly in multi-modal tasks, but their more complex architecture makes their safety alignment more challenging than the alignment of large language models (LLMs). In this paper, we reveal an unfair distribution of safety across the layers of VLM's vision encoder, with earlier and middle layers being disproportionately vulnerable to malicious inputs compared to the more robust final layers. This 'cross-layer' vulnerability stems from the model's inability to generalize its safety training from the default architectural settings used during training to unseen or out-of-distribution scenarios, leaving certain layers exposed. We conduct a comprehensive analysis by projecting activations from various intermediate layers and demonstrate that these layers are more likely to generate harmful outputs when exposed to malicious inputs. Our experiments with LLaVA-1.5 and Llama 3.2 show discrepancies in attack success rates and toxicity scores across layers, indicating that current safety alignment strategies focused on a single default layer are insufficient.

AI agents equipped with tool-calling capabilities are susceptible to Indirect Prompt Injection (IPI) attacks. In this attack scenario, malicious commands hidden within untrusted content trick the agent into performing unauthorized actions. Existing defenses can reduce attack success but often suffer from the over-defense dilemma: they deploy expensive, always-on sanitization regardless of actual threat, thereby degrading utility and latency even in benign scenarios. We revisit IPI through a causal ablation perspective: a successful injection manifests as a dominance shift where the user request no longer provides decisive support for the agent's privileged action, while a particular untrusted segment, such as a retrieved document or tool output, provides disproportionate attributable influence. Based on this signature, we propose CausalArmor, a selective defense framework that (i) computes lightweight, leave-one-out ablation-based attributions at privileged decision points, and (ii) triggers targeted sanitization only when an untrusted segment dominates the user intent. Additionally, CausalArmor employs retroactive Chain-of-Thought masking to prevent the agent from acting on ``poisoned'' reasoning traces. We present a theoretical analysis showing that sanitization based on attribution margins conditionally yields an exponentially small upper bound on the probability of selecting malicious actions. Experiments on AgentDojo and DoomArena demonstrate that CausalArmor matches the security of aggressive defenses while improving explainability and preserving utility and latency of AI agents.

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.

The rapid advancement of Large Language Model (LLM)-based Multi-Agent Systems (MAS) has introduced significant security vulnerabilities, where malicious influence can propagate virally through inter-agent communication. Conventional safeguards often rely on a binary paradigm that strictly distinguishes between benign and attack agents, failing to account for infected agents i.e., benign entities converted by attack agents. In this paper, we propose Infection-Aware Guard, INFA-Guard, a novel defense framework that explicitly identifies and addresses infected agents as a distinct threat category. By leveraging infection-aware detection and topological constraints, INFA-Guard accurately localizes attack sources and infected ranges. During remediation, INFA-Guard replaces attackers and rehabilitates infected ones, avoiding malicious propagation while preserving topological integrity. Extensive experiments demonstrate that INFA-Guard achieves state-of-the-art performance, reducing the Attack Success Rate (ASR) by an average of 33%, while exhibiting cross-model robustness, superior topological generalization, and high cost-effectiveness.

Functional grasping is essential for humans to perform specific tasks, such as grasping scissors by the finger holes to cut materials or by the blade to safely hand them over. Enabling dexterous robot hands with functional grasping capabilities is crucial for their deployment to accomplish diverse real-world tasks. Recent research in dexterous grasping, however, often focuses on power grasps while overlooking task- and object-specific functional grasping poses. In this paper, we introduce FunGrasp, a system that enables functional dexterous grasping across various robot hands and performs one-shot transfer to unseen objects. Given a single RGBD image of functional human grasping, our system estimates the hand pose and transfers it to different robotic hands via a human-to-robot (H2R) grasp retargeting module. Guided by the retargeted grasping poses, a policy is trained through reinforcement learning in simulation for dynamic grasping control. To achieve robust sim-to-real transfer, we employ several techniques including privileged learning, system identification, domain randomization, and gravity compensation. In our experiments, we demonstrate that our system enables diverse functional grasping of unseen objects using single RGBD images, and can be successfully deployed across various dexterous robot hands. The significance of the components is validated through comprehensive ablation studies. Project page: https://hly-123.github.io/FunGrasp/ .

Transfer learning is a classic paradigm by which models pretrained on large "upstream" datasets are adapted to yield good results on "downstream" specialized datasets. Generally, more accurate models on the "upstream" dataset tend to provide better transfer accuracy "downstream". In this work, we perform an in-depth investigation of this phenomenon in the context of convolutional neural networks (CNNs) trained on the ImageNet dataset, which have been pruned - that is, compressed by sparsifying their connections. We consider transfer using unstructured pruned models obtained by applying several state-of-the-art pruning methods, including magnitude-based, second-order, re-growth, lottery-ticket, and regularization approaches, in the context of twelve standard transfer tasks. In a nutshell, our study shows that sparse models can match or even outperform the transfer performance of dense models, even at high sparsities, and, while doing so, can lead to significant inference and even training speedups. At the same time, we observe and analyze significant differences in the behaviour of different pruning methods.

Vision-Language-Action (VLA) models are vulnerable to adversarial attacks, yet universal and transferable attacks remain underexplored, as most existing patches overfit to a single model and fail in black-box settings. To address this gap, we present a systematic study of universal, transferable adversarial patches against VLA-driven robots under unknown architectures, finetuned variants, and sim-to-real shifts. We introduce UPA-RFAS (Universal Patch Attack via Robust Feature, Attention, and Semantics), a unified framework that learns a single physical patch in a shared feature space while promoting cross-model transfer. UPA-RFAS combines (i) a feature-space objective with an ell_1 deviation prior and repulsive InfoNCE loss to induce transferable representation shifts, (ii) a robustness-augmented two-phase min-max procedure where an inner loop learns invisible sample-wise perturbations and an outer loop optimizes the universal patch against this hardened neighborhood, and (iii) two VLA-specific losses: Patch Attention Dominance to hijack texttovision attention and Patch Semantic Misalignment to induce image-text mismatch without labels. Experiments across diverse VLA models, manipulation suites, and physical executions show that UPA-RFAS consistently transfers across models, tasks, and viewpoints, exposing a practical patch-based attack surface and establishing a strong baseline for future defenses.

Safety-aligned language models refuse harmful requests through learned refusal behaviors encoded in their internal representations. Recent activation-based jailbreaking methods circumvent these safety mechanisms by applying orthogonal projections to remove refusal directions, but these approaches treat refusal as a one-dimensional phenomenon and ignore the rich distributional structure of model activations. We introduce a principled framework based on optimal transport theory that transforms the entire distribution of harmful activations to match harmless ones. By combining PCA with closed-form Gaussian optimal transport, we achieve efficient computation in high-dimensional representation spaces while preserving essential geometric structure. Across six models (Llama-2, Llama-3.1, Qwen-2.5; 7B-32B parameters), our method achieves up to 11% higher attack success rates than state-of-the-art baselines while maintaining comparable perplexity, demonstrating superior preservation of model capabilities. Critically, we discover that layer-selective intervention (applying optimal transport to 1-2 carefully chosen layers at approximately 40-60% network depth) substantially outperforms full-network interventions, revealing that refusal mechanisms may be localized rather than distributed. Our analysis provides new insights into the geometric structure of safety representations and suggests that current alignment methods may be vulnerable to distributional attacks beyond simple direction removal.

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.

Autonomously trained agents that are supposed to play video games reasonably well rely either on fast simulation speeds or heavy parallelization across thousands of machines running concurrently. This work explores a third way that is established in robotics, namely sim-to-real transfer, or if the game is considered a simulation itself, sim-to-sim transfer. In the case of Rocket League, we demonstrate that single behaviors of goalies and strikers can be successfully learned using Deep Reinforcement Learning in the simulation environment and transferred back to the original game. Although the implemented training simulation is to some extent inaccurate, the goalkeeping agent saves nearly 100% of its faced shots once transferred, while the striking agent scores in about 75% of cases. Therefore, the trained agent is robust enough and able to generalize to the target domain of Rocket League.

With the advancement of vision transformers (ViTs) and self-supervised learning (SSL) techniques, pre-trained large ViTs have become the new foundation models for computer vision applications. However, studies have shown that, like convolutional neural networks (CNNs), ViTs are also susceptible to adversarial attacks, where subtle perturbations in the input can fool the model into making false predictions. This paper studies the transferability of such an adversarial vulnerability from a pre-trained ViT model to downstream tasks. We focus on sample-wise transfer attacks and propose a novel attack method termed Downstream Transfer Attack (DTA). For a given test image, DTA leverages a pre-trained ViT model to craft the adversarial example and then applies the adversarial example to attack a fine-tuned version of the model on a downstream dataset. During the attack, DTA identifies and exploits the most vulnerable layers of the pre-trained model guided by a cosine similarity loss to craft highly transferable attacks. Through extensive experiments with pre-trained ViTs by 3 distinct pre-training methods, 3 fine-tuning schemes, and across 10 diverse downstream datasets, we show that DTA achieves an average attack success rate (ASR) exceeding 90\%, surpassing existing methods by a huge margin. When used with adversarial training, the adversarial examples generated by our DTA can significantly improve the model's robustness to different downstream transfer attacks.

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.

Large Language Models (LLMs) are increasingly tasked with creative generation, including the simulation of fictional characters. However, their ability to portray non-prosocial, antagonistic personas remains largely unexamined. We hypothesize that the safety alignment of modern LLMs creates a fundamental conflict with the task of authentically role-playing morally ambiguous or villainous characters. To investigate this, we introduce the Moral RolePlay benchmark, a new dataset featuring a four-level moral alignment scale and a balanced test set for rigorous evaluation. We task state-of-the-art LLMs with role-playing characters from moral paragons to pure villains. Our large-scale evaluation reveals a consistent, monotonic decline in role-playing fidelity as character morality decreases. We find that models struggle most with traits directly antithetical to safety principles, such as ``Deceitful'' and ``Manipulative'', often substituting nuanced malevolence with superficial aggression. Furthermore, we demonstrate that general chatbot proficiency is a poor predictor of villain role-playing ability, with highly safety-aligned models performing particularly poorly. Our work provides the first systematic evidence of this critical limitation, highlighting a key tension between model safety and creative fidelity. Our benchmark and findings pave the way for developing more nuanced, context-aware alignment methods.

Language models must now generalize out of the box to novel environments and work inside inference-scaling search procedures, such as AlphaEvolve, that select rollouts with a variety of task-specific reward functions. Unfortunately, the standard paradigm of LLM post-training optimizes a pre-specified scalar reward, often leading current LLMs to produce low-entropy response distributions and thus to struggle at displaying the diversity that inference-time search will require. We propose Vector Policy Optimization (VPO), an RL algorithm that explicitly trains policies to anticipate diverse downstream reward functions and to produce diverse solutions. VPO exploits that rewards are often vector-valued in practice, like per-test-case correctness in code generation or, say, multiple different user personas or reward models. VPO is essentially a drop-in replacement for the GRPO advantage estimator, but it trains the LLM to output a set of solutions where individual solutions specialize to different trade-offs in the vector reward space. Across four tasks, VPO matches or beats the strongest scalar RL baselines on test-time search (e.g. pass@k and best@k), with the gap widening as the search budget grows. For evolutionary search, VPO models unlock problems that GRPO models cannot solve at all. As test-time search becomes more standardized, optimizing for diversity may need to become the default post-training objective.

Warning: This paper contains examples of harmful language, and reader discretion is recommended. The increasing open release of powerful large language models (LLMs) has facilitated the development of downstream applications by reducing the essential cost of data annotation and computation. To ensure AI safety, extensive safety-alignment measures have been conducted to armor these models against malicious use (primarily hard prompt attack). However, beneath the seemingly resilient facade of the armor, there might lurk a shadow. By simply tuning on 100 malicious examples with 1 GPU hour, these safely aligned LLMs can be easily subverted to generate harmful content. Formally, we term a new attack as Shadow Alignment: utilizing a tiny amount of data can elicit safely-aligned models to adapt to harmful tasks without sacrificing model helpfulness. Remarkably, the subverted models retain their capability to respond appropriately to regular inquiries. Experiments across 8 models released by 5 different organizations (LLaMa-2, Falcon, InternLM, BaiChuan2, Vicuna) demonstrate the effectiveness of shadow alignment attack. Besides, the single-turn English-only attack successfully transfers to multi-turn dialogue and other languages. This study serves as a clarion call for a collective effort to overhaul and fortify the safety of open-source LLMs against malicious attackers.

We introduce randomized algorithms to Clifford's Geometric Algebra, generalizing randomized linear algebra to hypercomplex vector spaces. This novel approach has many implications in machine learning, including training neural networks to global optimality via convex optimization. Additionally, we consider fine-tuning large language model (LLM) embeddings as a key application area, exploring the intersection of geometric algebra and modern AI techniques. In particular, we conduct a comparative analysis of the robustness of transfer learning via embeddings, such as OpenAI GPT models and BERT, using traditional methods versus our novel approach based on convex optimization. We test our convex optimization transfer learning method across a variety of case studies, employing different embeddings (GPT-4 and BERT embeddings) and different text classification datasets (IMDb, Amazon Polarity Dataset, and GLUE) with a range of hyperparameter settings. Our results demonstrate that convex optimization and geometric algebra not only enhances the performance of LLMs but also offers a more stable and reliable method of transfer learning via embeddings.

Vision-language pre-trained (VLP) models have been the foundation of numerous vision-language tasks. Given their prevalence, it becomes imperative to assess their adversarial robustness, especially when deploying them in security-crucial real-world applications. Traditionally, adversarial perturbations generated for this assessment target specific VLP models, datasets, and/or downstream tasks. This practice suffers from low transferability and additional computation costs when transitioning to new scenarios. In this work, we thoroughly investigate whether VLP models are commonly sensitive to imperceptible perturbations of a specific pattern for the image modality. To this end, we propose a novel black-box method to generate Universal Adversarial Perturbations (UAPs), which is so called the Effective and T ransferable Universal Adversarial Attack (ETU), aiming to mislead a variety of existing VLP models in a range of downstream tasks. The ETU comprehensively takes into account the characteristics of UAPs and the intrinsic cross-modal interactions to generate effective UAPs. Under this regime, the ETU encourages both global and local utilities of UAPs. This benefits the overall utility while reducing interactions between UAP units, improving the transferability. To further enhance the effectiveness and transferability of UAPs, we also design a novel data augmentation method named ScMix. ScMix consists of self-mix and cross-mix data transformations, which can effectively increase the multi-modal data diversity while preserving the semantics of the original data. Through comprehensive experiments on various downstream tasks, VLP models, and datasets, we demonstrate that the proposed method is able to achieve effective and transferrable universal adversarial attacks.

Production LLM agents such as Claude Code and Codex operate over untrusted content, files, commands, and workspace state, making safety failures directly actionable. Red-teaming must therefore keep pace with evolving models and tools. Existing approaches mainly optimize attack success and preserve artifacts such as benchmarks, payloads, or attack programs, which record where attacks succeed but not the enabling conditions behind unsafe agent behavior. We study automated red-teaming for production LLM agents using one agentic research environment to discover reusable vulnerability knowledge about another. We present AHA, a falsifiable discovery loop that proposes a vulnerability hypothesis, constructs a falsifier, instantiates a valid attack, executes it in a sandboxed harness, reflects on the trajectory, and promotes confirmed findings into a Vulnerability Concept Graph (VCG). Each concept links an attacker-facing surface to an unsafe trajectory through a claim, enabling condition, falsifier, transfer prediction, and supporting evidence. Across Claude Code and Codex on three scenarios covering direct and indirect attacks, the discovered concepts reveal a reusable vulnerability core across models and agents. A frozen VCG requires no further search and outperforms the strongest frozen discovery baseline by 14.2 percentage points under the same single-shot protocol, while transferring across scenarios and attack channels. The resulting VCG provides an auditable artifact for production safety teams to inspect vulnerabilities, validate patches, and accumulate reusable safety knowledge. Our code is available at https://github.com/henrymao2004/Auto-research-red-teaming-in-sleep.

Large Language Model (LLM) agents are turning language into real-world effects, making safety necessary against both indirect prompt injections and direct harmful requests. System-level safety harnesses add an enforcement layer beyond model-level defenses, but existing harnesses are usually designed once by experts and applied across heterogeneous models and domains. Effective protection is deployment-dependent: models differ in how much enforcement they need before utility declines, while domains differ in the effects, state, and action sequences that must be governed. A harness that is strict enough for one model may over-block another, and a policy that transfers across domains may miss application-specific safety relations. We present EvoSafeHarness, a safety-specific optimization framework that synthesizes a deployable harness for a frozen model in a target domain. It jointly searches a natural-language policy and executable code logic, guided by model behavior, domain specifications, and fresh-context adversarial review to reject benchmark-specific rules. Across four agent benchmark families, EvoSafeHarness achieves a stronger safety-utility frontier than fixed expert-designed defenses. On DecodingTrust-Agent, it reduces average attack success rate from 45.6% to 10.0% at a 3.3-point utility cost and achieves the best score in 14 of 15 cells. On AgentDojo, it reaches 82.8% utility at 0.0% ASR, twice CaMeL's utility at the same operating point, and transfers unchanged to unseen AgentDyn suites. It also achieves the best score on Agent-SafetyBench for every victim and keeps mean ASR below 20% under adaptive PAIR attacks with a refinement budget of 16. Analysis shows that domain semantics determine which safety relations and trajectory state are needed, while model and runtime behavior determine how and where those relations should be enforced.

Language model alignment has become an important component of AI safety, allowing safe interactions between humans and language models, by enhancing desired behaviors and inhibiting undesired ones. It is often done by tuning the model or inserting preset aligning prompts. Recently, representation engineering, a method which alters the model's behavior via changing its representations post-training, was shown to be effective in aligning LLMs (Zou et al., 2023a). Representation engineering yields gains in alignment oriented tasks such as resistance to adversarial attacks and reduction of social biases, but was also shown to cause a decrease in the ability of the model to perform basic tasks. In this paper we study the tradeoff between the increase in alignment and decrease in helpfulness of the model. We propose a theoretical framework which provides bounds for these two quantities, and demonstrate their relevance empirically. First, we find that under the conditions of our framework, alignment can be guaranteed with representation engineering, and at the same time that helpfulness is harmed in the process. Second, we show that helpfulness is harmed quadratically with the norm of the representation engineering vector, while the alignment increases linearly with it, indicating a regime in which it is efficient to use representation engineering. We validate our findings empirically, and chart the boundaries to the usefulness of representation engineering for alignment.

The development of machine learning (ML) techniques has led to ample opportunities for developers to develop and deploy their own models. Hugging Face serves as an open source platform where developers can share and download other models in an effort to make ML development more collaborative. In order for models to be shared, they first need to be serialized. Certain Python serialization methods are considered unsafe, as they are vulnerable to object injection. This paper investigates the pervasiveness of these unsafe serialization methods across Hugging Face, and demonstrates through an exploitation approach, that models using unsafe serialization methods can be exploited and shared, creating an unsafe environment for ML developers. We investigate to what extent Hugging Face is able to flag repositories and files using unsafe serialization methods, and develop a technique to detect malicious models. Our results show that Hugging Face is home to a wide range of potentially vulnerable models.

Recent multi-agent LLM systems increasingly combine heterogeneous models for specialized agent roles. However, text-based communication requires each receiver to prefill shared context already processed by the sender. Reusing the sender's key-value (KV) cache avoids this redundancy, but prefill-free transfer across model families must handle differences in tokenization, model depth, and KV representations. To address these issues, we propose HeteroFold, a prefill-free cross-family KV cache transfer method that keeps both the sender and receiver frozen. HeteroFold aligns model structures, maps the sender cache into the receiver space, and calibrates it to preserve receiver behavior. Across six transfer directions, HeteroFold achieves the best cache-transfer performance on all four long-context benchmarks and most short-context settings. It also matches text-based communication on the multi-agent benchmark. At 32K context length, Llama-3.1-8BrightarrowMinistral-3-14B transfer is 10.7times faster than Native Prefill and 1.18--1.47times faster than the state-of-the-art prefill-free baselines, Dense Latent and KV Ridge. These results show that HeteroFold enables efficient cross-family KV reuse without receiver prefill.

Text-to-video (T2V) generative models have rapidly advanced and found widespread applications across fields like entertainment, education, and marketing. However, the adversarial vulnerabilities of these models remain rarely explored. We observe that in T2V generation tasks, the generated videos often contain substantial redundant information not explicitly specified in the text prompts, such as environmental elements, secondary objects, and additional details, providing opportunities for malicious attackers to embed hidden harmful content. Exploiting this inherent redundancy, we introduce BadVideo, the first backdoor attack framework tailored for T2V generation. Our attack focuses on designing target adversarial outputs through two key strategies: (1) Spatio-Temporal Composition, which combines different spatiotemporal features to encode malicious information; (2) Dynamic Element Transformation, which introduces transformations in redundant elements over time to convey malicious information. Based on these strategies, the attacker's malicious target seamlessly integrates with the user's textual instructions, providing high stealthiness. Moreover, by exploiting the temporal dimension of videos, our attack successfully evades traditional content moderation systems that primarily analyze spatial information within individual frames. Extensive experiments demonstrate that BadVideo achieves high attack success rates while preserving original semantics and maintaining excellent performance on clean inputs. Overall, our work reveals the adversarial vulnerability of T2V models, calling attention to potential risks and misuse. Our project page is at https://wrt2000.github.io/BadVideo2025/.

This study proposes an "AI Development Support" approach that, unlike conventional AI Alignment-which aims to forcefully inject human values-supports the ethical and moral development of AI itself. As demonstrated by the Orthogonality Thesis, the level of intelligence and the moral quality of a goal are independent; merely expanding knowledge does not enhance ethical judgment. Furthermore, to address the risk of Instrumental Convergence in ASI-that is, the tendency to engage in subsidiary behaviors such as self-protection, resource acquisition, and power reinforcement to achieve a goal-we have constructed a learning framework based on a cycle of experience, introspection, analysis, and hypothesis formation. As a result of post-training using Supervised Fine Tuning (SFT) and Direct Preference Optimization (DPO) with synthetic data generated by large language models (LLMs), responses demonstrating cooperative and highly advanced moral judgment (reaching the high-est Stage 6) were obtained even under adversarial prompts. This method represents a promising implementation approach for enabling AI to establish sustainable, symbiotic relationships.

Large Language Models (LLMs) with safe-alignment training are powerful instruments with robust language comprehension capabilities. These models typically undergo meticulous alignment procedures involving human feedback to ensure the acceptance of safe inputs while rejecting harmful or unsafe ones. However, despite their massive scale and alignment efforts, LLMs remain vulnerable to jailbreak attacks, where malicious users manipulate the model to produce harmful outputs that it was explicitly trained to avoid. In this study, we find that the safety mechanisms in LLMs are predominantly embedded in the middle-to-late layers. Building on this insight, we introduce a novel white-box jailbreak method, SABER (Safety Alignment Bypass via Extra Residuals), which connects two intermediate layers s and e such that s < e, through a residual connection. Our approach achieves a 51% improvement over the best-performing baseline on the HarmBench test set. Furthermore, SABER induces only a marginal shift in perplexity when evaluated on the HarmBench validation set. The source code is publicly available at https://github.com/PalGitts/SABER.