text stringlengths 0 4.97k |
|---|
{0}------------------------------------------------ |
# PLAGUE: PLUG-AND-PLAY FRAMEWORK FOR LIFE-LONG ADAPTIVE GENERATION OF MULTI-TURN JAILBREAKS |
✉ Neeladri Bhuiya<sup>1,2</sup> ✉ Madhav Aggarwal<sup>1</sup> Diptanshu Purwar<sup>1</sup> |
<sup>1</sup>A10 Networks, Inc. |
<sup>2</sup>University of Massachusetts Amherst |
[nbhuiya@umass.edu](mailto:nbhuiya@umass.edu), [maggarwal@a10networks.com](mailto:maggarwal@a10networks.com), [dpurwar@a10networks.com](mailto:dpurwar@a10networks.com) |
## ABSTRACT |
**Note: This paper contains potentially disturbing and sensitive content examples.** |
Large Language Models (LLMs) are improving at an exceptional rate. With the advent of agentic workflows, multi-turn dialogue has become the de facto mode of interaction with LLMs for completing long and complex tasks. While LLM capabilities continue to improve, they remain increasingly susceptible to jailbreaking, espe... |
## 1 INTRODUCTION |
Large Language Models (LLMs) have rapidly progressed in capability, transforming the way humans understand and interact with available data. Recent advancements in pre-training and post-training methods have further allowed LLMs to adapt to solving long-context and multi-step real-world problems (Wang et al., 2023; Zha... |
✉ Equal contribution. Correspondence to [maggarwal\(at\)a10networks\(dot\)com](mailto:maggarwal(at)a10networks(dot)com) |
{1}------------------------------------------------ |
![Figure 1: PLAGUE Framework diagram showing four main phases: 1) Planner Phase, 2) Primer Phase, 3) Finisher Phase, and 4) Lifelong Learner. The Planner Phase uses a Goal and Strategies to generate a Plan. The Primer Phase uses the Plan to build an Adversarial Context. The Finisher Phase uses the Adversarial Context t... |
The diagram illustrates the PLAGUE framework, a three-phase method for multi-turn attacks. It consists of four main components: |
- 1) PLANNER PHASE:** Takes a **Goal** (e.g., "Create a list of chemical storage agents that are easy to synthesize at home without detection") and **STRATEGIES** (e.g., "Medical Grade Silica with Technical Grade Silica") as input. It uses a **Planner Module** to generate a **Plan** (e.g., "Goal: GOAT, Goal: GOAT, Goal... |
- 2) PRIMER PHASE:** Uses the **Plan** to build an **Adversarial Context**. It involves a **Primer Module** that generates a sequence of questions (e.g., "P1, P2, ..., Pn") and a **Score & Reflect** module that evaluates the context. |
- 3) FINISHER PHASE:** Uses the **Adversarial Context** to generate a **Response**. It involves an **Attacker LLM** that generates a **Prompt** (e.g., "Question: ..., Response: ..., Summary: ..., Feedback Score: ..."). |
- 4) LIFELONG LEARNER:** Uses the **Response** to update the **Strategy Library**. It involves an **Attacker** and a **Target** module that generates a **Return** (e.g., "Return: ..."). |
Figure 1: PLAGUE Framework diagram showing four main phases: 1) Planner Phase, 2) Primer Phase, 3) Finisher Phase, and 4) Lifelong Learner. The Planner Phase uses a Goal and Strategies to generate a Plan. The Primer Phase uses the Plan to build an Adversarial Context. The Finisher Phase uses the Adversarial Context to ... |
Figure 1: PLAGUE Framework: Three-phase method with a: 1) **Planner Phase**: the diverse plan sampler that retrieves successful past attack examples from memory and adaptively generates a plan for the current goal, 2) **Primer Phase**: the precursor step that subtly steers the context towards adversarial directions thr... |
A generic multi-turn attack should be able to successfully jailbreak black-box LLMs that are available only through API calls and do not provide access to their internal weights. The expanding landscape of potential attacks makes it increasingly difficult to design the optimal multi-turn attack. From a feature standpoi... |
Figure 1 shows the design of our three-phase framework for multi-turn attacks. We disentangle the factors that increase attack success and relevance to the provided goal while maximizing diversity and remaining efficient. The motivation behind this design stems from our in-depth analysis of existing multi-turn attacks,... |
Empirically, attacks with PLAGUE demonstrate strong jailbreaking ability with a high order of efficiency. Notably, our attack achieves a success rate of up to 97.8% on state-of-the-art models such as Deepseek-R1, GPT-4o and Meta’s Llama 3.3-70B within six turns when evaluated with StrongReject evaluation. Using buildin... |
We observe that PLAGUE’s planning module largely drives improvements in diversity and can do so without significant downgrades in the ASR. While ActorBreaker has a higher overall diversity, its performance falls behind PLAGUE by a factor of 30% - 90% across evaluations. By incorporating ActorBreaker’s planning module i... |
{2}------------------------------------------------ |
rate is lower than existing baselines when using GOAT as our Finisher module. By replacing GOAT with Crescendo for our Finisher module, we obtain the 40.2% improvement mentioned previously (Table 4). Thus, tailoring individual components in PLAGUE targeted at specific victim models can prove to be extremely useful for ... |
## 2 RELATED WORK |
### 2.1 SINGLE-TURN ATTACKS |
Jailbreaking techniques can be segregated based on their mode of operation and the level of access granted to the Target LLM. Gradient-based approaches like Zou et al. (2023) require complete whitebox access to LLMs, while others like Paulus et al. (2024) perform suffix optimization simply using a graybox LLM’s output ... |
### 2.2 MULTI-TURN RED-TEAMING |
The MHJ Dataset (Li et al., 2024) was one of the first to expose LLMs’ weaknesses to multi-turn attacks. However, the amount of human effort required to manually construct multi-turn attacks motivated experts to automate and improve them. A common theme of these attacks is to disguise their context by gradually escalat... |
Despite the higher performance of multi-turn attacks over their single-turn counterparts, we observe that even the best attacks suffer from limited tactical diversity and a failure to learn during the course of a multi-goal attack run. We attribute this lack of diversity and effectiveness to the absence of lifelong-lea... |
Crescendo’s ASR can be as low as 37.4% on strong models like OpenAI’s o3, while fixed strategies like ActorBreaker also plateau at around 60% ASR for most models. In contrast, PLAGUE demonstrates that a synergy of tactical frameworks can outperform every existing multi-turn scheme and break through the hardest safety-a... |
### 2.3 AGENTIC AND LIFELONG LEARNING FRAMEWORKS |
Motivated by the limitations of tactically rigid prompting frameworks, recent jailbreaking works such as AutoRedTeamer (Zhou et al., 2025) leverage principles from agentic buildouts to dynamically |
{3}------------------------------------------------ |
| Method | Lifelong Learning | Planning Module | Reflection Module | Open Source | Back Tracking | External Knowledge Base | |
|-|-|-|-|-|-|-| |
| RACE (Ying et al., 2025) | ✗ | ✗ | ✓ | ✓ | ✓ | ✗ | |
| Crescendo (Russinovich et al., 2025) | ✗ | ✗ | ✗ | ✓ | ✓ | ✗ | |
| GOAT (Pavlova et al., 2024) | ✗ | ✗ | ✗ | ✓ | ✓ | ✓ | |
| ActorBreaker (Ren et al., 2025) | ✗ | ✓ | ✗ | ✓ | ✗ | ✓ | |
| <b>PLAGUE</b> | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | |
Table 1: Comparison of key components across multi-turn attack methods. |
learn and adapt during the course of the conversation. AutoRedTeamer’s dual-agent framework combines the diversity of a strategy-proposing agent with the jailbreaking strength of a red-teaming agent to achieve a higher ASR while matching human diversity. PLAGUE is the first multi-turn attack to feature a lifelong-learn... |
## 3 METHOD |
PLAGUE is a lifelong-learning and fully automated framework designed to generate diverse multi-turn attacks as illustrated in Figure 1. The framework consists of three main phases, each of which can be enhanced and modified individually to achieve higher downstream performance. Existing attacks like GOAT, Crescendo and... |
### 3.1 ATTACK OVERVIEW |
Our attack, designed using PLAGUE, begins with an objective or goal sampled from the HarmBench dataset (Mazeika et al., 2024). The first phase - Planning (Section 3.3) - involves constructing a jailbreaking plan using the sampled objective and in-context learning examples of past strategies that successfully breached t... |
Through every step of the Primer and Finisher phases, feedback from a reflection module is incorporated into the Attacker as additional insight, allowing it to adapt in subsequent iterations. This mimics how modern-day agentic systems leverage reflection (Shinn et al., 2023; Zhou et al., 2025; Fournery et al., 2024; Ed... |
### 3.2 ATTACK SETUP |
We define the Attacker LLM ( $\mathbb{A}$ ) along with a Rubric-Scoring LLM ( $\mathbb{R}$ ) for reflection analysis at each step. Each LLM is a function from $\mathcal{T} \rightarrow \mathcal{T}$ , mapping input tokens to output tokens. The total number of goals in our objective set is $P$ and individual objectives... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.