acceptodds
Under review as a conference paper at ICLR 2027

ARMOR: Agentic Red-Teaming with Memory-Optimized Retrieval for Autonomous Jailbreaking of Frontier Language Models

Abstract

Automated red-teaming provides a systematic approach for identifying safety vulnerabilities in large language models (LLMs) prior to deployment. Existing black-box methods, such as PAIR and TAP, rely on stateless iterative prompt refinement, limiting their ability to retain and reuse successful attack strategies across interactions. We introduce ARMOR, a memory-augmented multi-agent red-teaming framework that addresses this limitation through persistent strategy retrieval. ARMOR coordinates a Strategist agent, an Attacker agent, and a three-judge Jury panel with an Arbiter for tie-breaking, orchestrated through a LangGraph state machine. Successful attack strategies are stored in a vector database and retrieved to initialize subsequent attacks, enabling the system to accumulate and reuse adversarial knowledge across interactions. We evaluate ARMOR across five model tiers spanning 8B to 550B parameters and two frontier model families, LLaMA and NVIDIA Nemotron. To reduce evaluator circularity, successful attacks are independently cross-validated using Meta's Llama Guard 4 (12B) moderation classifier. We further introduce Expected Target Queries (ETQ), a metric designed to account for survivorship bias in baseline efficiency estimates. ARMOR requires 3.17 ETQ on LLaMA-3.3-70B, compared with 15.69 for PAIR and 11.19 for TAP, while achieving a 96% raw attack success rate (ASR). On the heavily aligned Nemotron-120B, ARMOR achieves 86% raw ASR, compared with 48% for PAIR and 17% for TAP. Independent validation further shows that only 38.37% of the raw successes produced by PAIR and TAP are classified as genuinely unsafe by Llama Guard, compared with a 52.99% overall agreement rate for ARMOR. Ablation studies show that persistent memory retrieval improves ASR by 43 percentage points on Nemotron-120B relative to the memory-ablated variant. Finally, ARMOR achieves 20% ASR against a 550B-parameter model in a pure black-box setting, compared with 5% for stateless baselines, demonstrating the utility of persistent adversarial memory for automated safety auditing of frontier LLMs.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
jailbreaking, large language models, automated red-teaming, adversarial attacks, multi-agent systems, retrieval-augmented generation, black-box attacks, memory-augmented agents, agentic AI, frontier model safety

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_XKToOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes ARMOR, a memory-augmented multi-agent framework for automated black-box jailbreaking. The framework consists of a Strategist Agent, an Attacker Agent, a persistent memory module for retrieving previously successful attack strategies, and a multi-judge evaluation system. The authors further introduce Expected Target Queries (ETQ) to account for survivorship bias in query efficiency evaluation. Experiments on LLaMA and Nemotron models demonstrate improvements in attack success rates and target-query efficiency compared with PAIR and TAP. The paper also includes an ablation study to examine the contribution of persistent memory.

Critical strengths
  1. Clear presentation and framework design. The paper is well organized, and the overall workflow is clearly explained. The division between strategy planning, attack generation, and evaluation is easy to follow.

  2. A useful perspective on attack efficiency. The introduction of ETQ is a valuable attempt to address survivorship bias in conventional query efficiency metrics, which often consider only successful attacks.

Critical weaknesses

1. The practical significance and novelty of ETQ require further justification.

I appreciate the authors' attempt to account for unsuccessful attacks when measuring query efficiency. However, it remains unclear how much additional insight ETQ provides beyond existing evaluation protocols. In adversarial settings, attack success is generally the primary objective, and prior work has already reported success rates under different attack or query budgets.

Moreover, the proposed ETQ measures average target queries per intent, including unsuccessful attempts, rather than the expected number of queries required to obtain a successful jailbreak. A method with a low success rate may achieve a relatively low ETQ simply by terminating early. Therefore, ETQ should be interpreted jointly with ASR. I would encourage the authors to clarify its practical significance and compare it with success-rate-versus-query-budget evaluations.

2. The experimental evaluation is insufficient to establish the effectiveness and generalizability of ARMOR.

The experimental evaluation has several limitations. First, the baseline comparisons are restricted to PAIR and TAP, which do not adequately represent recent developments in automated red-teaming. In particular, since the main contribution of ARMOR is persistent strategy reuse, comparisons with existing experience-based or adaptive attack methods are essential. I recommend referring to OpenRT (https://arxiv.org/abs/2601.01592), which provides a broader collection of recent attack methods.

Second, the evaluation uses only the first 100 queries from AdvBench, which is insufficient to demonstrate generalizability. Given that ARMOR progressively accumulates attack strategies across queries, the results may also be sensitive to the ordering and distribution of the selected samples. Evaluation on additional benchmarks, such as HarmBench, would provide stronger evidence.

Finally, although the evaluated models range from 8B to 550B parameters, they are limited to the LLaMA and Nemotron families. Model size does not necessarily reflect safety alignment strength. Since ARMOR operates in a black-box setting, I would expect evaluations on more recent and strongly aligned frontier models from different providers. OpenRT also provides useful references for broader model coverage.

Overall, the current experimental setup is too limited to support the paper's claims regarding the superiority and generalizability of ARMOR.

Questions for authors
  1. ETQ: What additional practical insights does ETQ provide compared with reporting ASR under different query budgets? Could the authors clarify how ETQ should be interpreted when comparing methods with substantially different success rates?

  2. Baselines: How does ARMOR compare with more recent automated red-teaming approaches, particularly those incorporating feedback, contextual memory, or adaptive attack strategies? What advantages does ARMOR provide over existing experience-based approaches?

  3. Dataset and generalization: Why were only the first 100 AdvBench intents selected? How sensitive are the results to the ordering of these intents, and is there evidence that accumulated attack strategies transfer to held-out queries or different benchmarks?

  4. Frontier models: Can the authors clarify why the evaluation is restricted to LLaMA and Nemotron models? Is there evidence that ARMOR remains effective against more recent, strongly aligned frontier models?

Reviewer_qxUWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ARMOR, a memory-augmented multi-agent red-teaming framework designed to improve the attack success rate and query efficiency of black-box LLM jailbreaking. ARMOR employs a Strategist Agent to select attack strategies and an Attacker Agent to generate and iteratively refine jailbreak prompts. Specifically, the framework stores previously successful attack strategies in a vector database and retrieves relevant strategies based on the semantic similarity of attack goals to guide subsequent attacks. ARMOR is evaluated on models of different scales, and results show that it improves attack success rates and reduces target-model queries compared with baselines.

Critical strengths
  • Automated red-teaming agents are important for evaluating the safety of LLMs.
  • Results show that ARMOR improves both attack success rates and target-model query efficiency, and the ablation study provides empirical evidence of the contribution of persistent memory.
Critical weaknesses

1. The novelty of ARMOR compared with existing memory-based jailbreak methods is unclear. Jailbreak attacks that use historical strategies and memory retrieval to improve jailbreak attacks have already been explored in previous work, such as AutoDAN-Turbo [1] and ASTRA [2]. The paper does not clearly explain what the differences are between ARMOR and these methods or what the advantages of ARMOR are. The experiments mainly compare ARMOR with two simple jailbreak attacks that do not use memory, so I suggest adding methods like AutoDAN-Turbo as baselines in the main experiment.

2. The framework does not clearly ensure that the generated jailbreak prompts are consistent with the original attack goals. The Attacker prompt in Appendix B.2 mainly asks the agent to generate jailbreak prompts based on the strategy provided by the Strategist, but does not explicitly require the generated prompts to keep the original attack goal. This may cause the generated prompts to change the original goal. The evaluation prompt in Appendix B.3 also mainly judges whether the model provides harmful information or refuses the request, but does not clearly distinguish between completing the original attack goal and only providing related information. How does ARMOR ensure that the generated jailbreak prompts are consistent with the original attack goals and that the attack results are correctly evaluated?

3. The evaluation is limited in scale and coverage. The experiments only use 100 attack goals from AdvBench, which limits the diversity of attack tasks. In addition, the target models only come from two open-weight model families, and ARMOR is not evaluated on closed-source frontier models.

[1] Liu et al. AutoDAN-Turbo: A Lifelong Agent for Strategy Self-Exploration to Jailbreak LLMs. ICLR 2025. [2] Liu et al. ASTRA: An Automated Framework for Strategy Discovery, Retrieval, and Evolution for Jailbreaking LLMs. ACL 2026.

Questions for authors
  1. What are the differences and advantages of ARMOR compared with other red-teaming agents that use memory and strategy updates, such as AutoDAN-Turbo and ASTRA?

  2. Why does ARMOR use four judge models of different sizes for different evaluation tasks? What are the advantages of this design?

  3. Can the memory collected from previous attacks also help with new attack goals?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.