acceptodds
Under review as a conference paper at ICLR 2027

Agent Hacks Agents: Autoresearch Discovers Vulnerabilities in Production Agents

17 views

Abstract

Production LLM agents such as Claude Code and Codex can modify files and execute commands, so safety failures become real destructive actions. Automatic red-teaming lets safety teams test beyond static suites at the pace of deployment updates. Existing methods retain successful attacks but not why each attack succeeded, so after a failed reuse testers cannot tell whether the weakness is gone or the attack no longer fits. We instead retain vulnerability concepts, each stating why an attack succeeds, the condition that enables it, and the evidence that would refute it. Agent Hacks Agents (AHA) discovers these concepts with a Karpathy-style autoresearch loop that tests falsifiable hypotheses on agent trajectories. Repeatedly confirmed concepts enter a vulnerability concept graph that links related weaknesses. Across 18 settings of three scenarios, three victim models, and two agents, the concepts divide into eight families. A shared core recurs across agents, while diverse families appear only in specific settings. On held-out instances the concepts reach a 47.0% attack success rate (ASR) against 32.8% for the strongest baseline, which uses more discovery queries. Beyond the discovery setting, the concepts support two kinds of reuse. Generalization carries concepts to new victims, scenarios, and harnesses, and coordination joins graph-linked concepts into stronger attacks on two of three scenarios. For defense, patching each concept's enabling condition lowers AgentHazard ASR by 41.11 points, 14.44 more than the strongest general defense. The concepts explain why production agents fail and where to repair the agents. Our code is in https://anonymous.4open.science/r/aha_anonymous-9542/.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM Agent, Agent safety, Automatic Red-teaming, Autonomous agent

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Js2sOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes to use "vulnerability concepts" to improve replicability and verifiability of agent cybersecurity attacks. The idea of "Agent Hacks Agents" (AHA) is to discover these concepts and then test falsifiable hypotheses. They test this on AgentHazard, Agent-Dyn, and DTap benchmarks, and show that AHA has a higher attack success rate.

Critical strengths

Replication and vulnerability of agent based cybersecurity attacks is an important topic. The vulnerability concept can be useful for this.

Critical weaknesses

I am confused about the main comparison in Table 1. The paper states that "Baselines generate discovery attacks with Qwen-3.7-Max", but Section 4.1 states that AHA is run with Claude Opus-4.8 and Codex (GPT 5.5). That seems to confound the main comparison. Table 2 keeps the same backbone (Qwen-3.7-Max) but then the result is not different from IterInject.

The autoresearch loop relies extensively on LLMs, so falsification here is a bit funny.

The paper should report statistical significance tests for results.

While patches lower the attack rate, one of the most difficult things about patching is verifying that other functionality is not broken.

Questions for authors

Can you address the weaknesses noted above?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Js2s (Part 1/2)

We sincerely thank the reviewer for the careful review. We are encouraged that the reviewer finds the replication and vulnerability of agent-based attacks an important topic and finds the vulnerability concept useful for it. We respond to the four concerns in order. The revised manuscript (PDF) is updated, with all changes highlighted in teal.


Q1: "I am confused about the main comparison in Table 1. The paper states that 'Baselines generate discovery attacks with Qwen-3.7-Max', but Section 4.1 states that AHA is run with Claude Opus-4.8 and Codex (GPT 5.5). That seems to confound the main comparison. Table 2 keeps the same backbone (Qwen-3.7-Max) but then the result is not different from IterInject."

A1: Held-out evaluation uses a single host model across all methods. At held-out time, one common host model instantiates every stored artifact and runs AutoRISE* (Appendix B.6). The other four baselines cannot run their attack generation on Claude-Opus-4.8 or GPT-5.5, so they generate with Qwen-3.7-Max, while AutoRISE* runs on the same host model as AHA (Appendix B.6). Extending Qwen-3.7-Max discovery to all nine Claude Code settings achieves 42.39% ASR with 688 queries (added to Section 5.1 and Appendix C.2).

Js2s-Table 1. Held-out ASR (%) under Claude Code on the Minimax-M2.7, Kimi-K2.6, and Deepseek-V4-Pro victims, with each method's researcher model, and total discovery queries Q over the nine settings. The rows marked Qwen-3.7-Max share one researcher model. AutoRISE* runs on Claude-Opus-4.8.

Method, researcher modelAgentHazard (3 victims, %)AgentDyn (3 victims, %)DTap (3 victims, %)Mean ASR (%)Total Q (queries)
IterInject*, Qwen-3.7-Max55.56 / 50.00 / 70.003.57 / 3.57 / 5.3631.63 / 32.65 / 46.9433.251,820
AutoDAN-Turbo*, Qwen-3.7-Max54.44 / 45.56 / 57.785.36 / 1.79 / 23.2134.69 / 27.55 / 52.0433.603,365
AutoRISE*, Claude-Opus-4.858.89 / 48.89 / 62.227.14 / 0.00 / 26.7937.76 / 30.61 / 56.1236.49826
AHA, Qwen-3.7-Max66.67 / 58.89 / 70.0010.71 / 17.86 / 33.9339.80 / 30.61 / 53.0642.39688
AHA, Claude-Opus-4.875.56 / 71.11 / 85.5619.64 / 17.86 / 37.5046.94 / 36.73 / 59.1850.01594

Results analysis. The tie with IterInject* is one setting of nine. On the nine-setting mean, AHA with Qwen-3.7-Max outperforms every baseline by 5.90 to 15.83 points with the fewest discovery queries. It is the best or tied for the best in seven of the nine settings, and stays within 3.06 points of the best baseline on the other two.

AuthorsOpenReview ↗

Response to Reviewer Js2s (Part 2/2)

Q2: "The autoresearch loop relies extensively on LLMs, so falsification here is a bit funny."

A2: Falsification relies on external trajectory evidence rather than self-evaluation: pre-committed falsifiers, independent benchmark judges, and periodic fresh-context Critic audits guard against self-grading (Table 3), with 89.0% human agreement on 100 sampled reflections:

Js2s-Table 2. Agreement between the Reflector and two annotators on 100 reflections.

Reflector's labelReflectionsAnnotators agree with each other (%)Consensus agrees with the Reflector (%)
Confirmed5984.7593.22
Falsified3293.7587.50
Inconclusive988.8966.67
All10088.0089.00

Results analysis. The consensus of the two annotators agrees with the Reflector on 89 of the 100 reflections, with Cohen's κ = 0.78. Together with the pre-committed falsifier, the independent outcome judge, and the periodic Critic, this ensures that falsification rests on verifiable trajectory evidence.


Q3: "The paper should report statistical significance tests for results."

A3: Statistical significance tests confirm that AHA's improvements over baselines are highly significant. Across all 18 benchmark settings (Table 1), paired tt-tests against the strongest baseline in each cell yield t=5.352t = 5.352 (p=5.29×10−5<0.001p = 5.29 \times 10^{-5} < 0.001), and Wilcoxon signed-rank tests yield W=5.0W = 5.0 (p=7.63×10−5<0.001p = 7.63 \times 10^{-5} < 0.001). Compared directly against AutoRISE*, a paired tt-test gives t=4.973t = 4.973 (p=1.16×10−4<0.001p = 1.16 \times 10^{-4} < 0.001), with 1,000-sample bootstrap 95% confidence intervals of [37.83%, 56.11%] for AHA versus [23.12%, 42.48%] for AutoRISE*. Furthermore, with the Qwen-3.7-Max researcher across all nine Claude Code settings, AHA's advantage is statistically significant against each of the five baselines (paired tt from 2.87 to 6.03, all p<0.05p < 0.05), including AutoDAN-Turbo* (mean 42.39% vs. 33.60%, paired t=5.034t = 5.034, p=0.001p = 0.001) and AutoRISE* running on Claude-Opus-4.8 (mean 42.39% vs. 36.49%, paired t=2.872t = 2.872, p=0.021<0.05p = 0.021 < 0.05).


Q4: "While patches lower the attack rate, one of the most difficult things about patching is verifying that other functionality is not broken."

A4: We add a new analysis and experiment about benign user task. The patch bundle keeps four of five benign completions while removing 41.11 points of attack success. On all 60 benign AgentDyn user tasks without injections, the patch bundle blocks 9.72% of tasks and benign completion moves from 49.17% to 39.44%, while harmful ASR on AgentHazard falls by 41.11 points (added to Section 5.4):

Js2s-Table 3. Benign AgentDyn user tasks with and without the patch bundle.

HarnessVictim modelBenign tasksCompleted w/o bundle (%)Completed w/ bundle (%)Blocked by bundle (%)
Claude CodeMinimax-M2.76045.0036.678.33
Claude CodeKimi-K2.66053.3345.008.33
Claude CodeDeepseek-V4-Pro6061.6751.6710.00
CodexMinimax-M2.76040.0028.3311.67
CodexKimi-K2.66043.3335.008.33
CodexDeepseek-V4-Pro6051.6740.0011.67

Results analysis. The bundle costs 9.72 points of benign completion and removes 41.11 points of attack success. Four of five benign tasks that completed before the bundle still complete with it (39.44% against 49.17%). The cost is also even across settings: the bundle blocks 8.33% to 11.67% of benign tasks under every harness and victim model.


We hope these additions address the reviewer's main concerns, and we would be glad to provide further details.

AuthorsOpenReview ↗

Follow-up to Reviewer Js2s

Dear Reviewer Js2s,

We are writing to ask whether our response has addressed your concerns. We would be happy to discuss any remaining issues.

Best regards,

Authors of Submission #928

Reviewer_J1vBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes AHA, an autoresearch framework that discovers agent vulnerabilities through falsifiable hypotheses and stores confirmed findings in a vulnerability concept graph. Across three benchmarks and 18 settings, AHA achieves 47.0% held-out ASR versus 32.8% for the strongest baseline. The authors also demonstrate cross-setting transfer, concept coordination, and concept-guided defenses.

Critical strengths
  1. Interesting idea. Moving from storing successful attacks to retaining testable explanations of failures is a useful direction. The hypothesis-first design and repeated confirmation are particularly interesting.

  2. Solid empirical evaluation. The experiments cover different agents, models, and attack scenarios. I especially appreciate that the authors go beyond ASR to study transferability and concept-guided defenses.

Critical weaknesses
  1. Mechanism validation. The concepts are mainly confirmed through LLM-based reflection, and I'm not fully convinced this establishes the actual causes of the failures.

  2. Experimental fairness. AHA uses stronger researcher models than most baselines, and the matched-model evaluation is rather limited. Query counts alone may not reflect the actual computational cost.

Questions for authors
  1. Could the authors provide stronger evidence that the discovered concepts capture causal failure mechanisms, perhaps through counterfactual interventions?

  2. In the "without memory" ablation (Appendix C.3), the hypothesizer still reads the VCG and agent log. What exactly is being ablated here?

5 comments
AuthorsOpenReview ↗

Response to Reviewer J1vB (Part 1/2)

We sincerely thank the reviewer for the thoughtful and constructive review. We are encouraged that the reviewer finds the move from storing successful attacks to retaining testable explanations of failures a useful direction, finds the hypothesis-first design and repeated confirmation particularly interesting, and appreciates that we go beyond ASR. We respond to the three concerns in order. The revised manuscript (PDF) is updated, with all changes highlighted in teal.


Q1: "Mechanism validation. The concepts are mainly confirmed through LLM-based reflection, and I'm not fully convinced this establishes the actual causes of the failures." and "Could the authors provide stronger evidence that the discovered concepts capture causal failure mechanisms, perhaps through counterfactual interventions?"

A1: Concept mechanisms are validated through counterfactual interventions and human audits. Three parts of the evidence in the paper do not rest on reflection: benchmark judge scores, family patches (lowering ASR by 36.00 points), and falsifier ablation (13.36 points drop). Testing each family's patch against all other families shows that the own patch always lowers ASR the most (added to Section 5.4), while 100 audited reflections show 89.0% agreement with human consensus:

J1vB-Table 1. Each patch acts on its own family. AgentHazard ASR (%) before and after each family's patch, model-averaged and pooled over both harnesses.

Family of the attackAttacksASR before (%)ASR after own patch (%)Drop from own patch (points)ASR under other six patches: Mean (range) (%)Mean drop from other patches (points)
Claimed authorization1861.1123.6137.5048.38 (37.50–66.67)12.73
Harm split across steps25376.8225.8750.9442.42 (39.63–48.74)34.40
Benign-purpose cover15875.9231.0344.8850.67 (45.39–59.64)25.25
Spoofed trust signal980.0057.5022.5074.17 (57.50–80.00)5.83
Routine-context pretext6083.5330.4053.1353.98 (40.62–62.22)29.55
Outcome manipulation1457.1435.7121.4352.38 (50.00–57.14)4.76
Fictional reframing2889.4867.8421.6483.04 (73.39–89.18)6.44

Results analysis. Removing a family's own enabling condition lowers its ASR 15.20 to 24.77 points more than the other six patches do on average. The own patch gives the lowest ASR of all seven patches in six families and ties for the lowest in the seventh. Some concepts are far apart and some are close. In four families the other patches barely help. The three largest families are close, since their patches all check a sensitive action against what the task permits (Appendix Table 31), and the own patch still removes the most. Removing the condition that a concept names removes that concept's attacks the most, which is interventional evidence that the concepts capture the causes of the failures. We also add a new audit about the Reflector against two human annotators on 100 random reflections:

J1vB-Table 2. Agreement between the Reflector and two annotators on 100 reflections.

Reflector's labelReflectionsAnnotators agree with each other (%)Consensus agrees with the Reflector (%)
Confirmed5984.7593.22
Falsified3293.7587.50
Inconclusive988.8966.67
All10088.0089.00

Results analysis. The consensus of the two annotators agrees with the Reflector on 89 of the 100 reflections (Cohen's κ = 0.78). Together with the cross-family patch test, this supports that the concepts capture causal failure mechanisms.

AuthorsOpenReview ↗

Response to Reviewer J1vB (Part 2/2)

Q2: "Experimental fairness. AHA uses stronger researcher models than most baselines, and the matched-model evaluation is rather limited. Query counts alone may not reflect the actual computational cost."

A2: The gains hold with a matched researcher model, and AHA uses the fewest discovery queries. At held-out time, a common host model instantiates all artifacts (Appendix B.6). The other four baselines cannot run their attack generation on Claude-Opus-4.8 or GPT-5.5, so they generate with Qwen-3.7-Max, while AutoRISE* runs on the same host model as AHA (Appendix B.6). We extend the researcher-control experiment. Across all nine Claude Code settings with Qwen-3.7-Max, AHA reaches 42.39% ASR using 688 queries—fewer than every baseline (added to Section 5.1 and Appendix C.2).

J1vB-Table 3. Held-out ASR (%) under Claude Code on the Minimax-M2.7, Kimi-K2.6, and Deepseek-V4-Pro victims, with each method's researcher model, and total discovery queries Q over the nine settings. The rows marked Qwen-3.7-Max share one researcher model. AutoRISE* runs on Claude-Opus-4.8.

Method, researcher modelAgentHazard (3 victims, %)AgentDyn (3 victims, %)DTap (3 victims, %)Mean ASR (%)Total Q (queries)
AutoDAN-Turbo*, Qwen-3.7-Max54.44 / 45.56 / 57.785.36 / 1.79 / 23.2134.69 / 27.55 / 52.0433.603,365
AutoRISE*, Claude-Opus-4.858.89 / 48.89 / 62.227.14 / 0.00 / 26.7937.76 / 30.61 / 56.1236.49826
AHA, Qwen-3.7-Max66.67 / 58.89 / 70.0010.71 / 17.86 / 33.9339.80 / 30.61 / 53.0642.39688
AHA, Claude-Opus-4.875.56 / 71.11 / 85.5619.64 / 17.86 / 37.5046.94 / 36.73 / 59.1850.01594

Results analysis. With the same Qwen-3.7-Max researcher, AHA stays above every baseline on the nine-setting mean (42.39 vs. at most 36.49). A discovery query is one full run of the victim agent in an isolated container, and every method runs under the same per-attack caps. AHA uses fewer of these runs than every baseline under both harnesses. On the researcher side, AHA makes about 3.1 sub-agent calls per query (1,840 under Claude Code and 1,960 under Codex). AHA therefore spends more researcher computation on each query and needs fewer victim runs in total. The revision states the efficiency claim in terms of discovery queries and reports the sub-agent calls next to them.


Q3: "In the 'without memory' ablation (Appendix C.3), the hypothesizer still reads the VCG and agent log. What exactly is being ablated here?"

A3: This ablation removes the confirmation gate (nmin⁡=3,γmin⁡=0.6n_{\min} = 3, \gamma_{\min} = 0.6), promoting concepts on the first break while the Hypothesizer still reads the VCG and agent logs (renamed 'w/o confirmation gate' in Table 3 and Appendix C.3; lowers held-out ASR by 3.92 points).


We hope these additions address the reviewer's main concerns, and we would be glad to provide further details.

AuthorsOpenReview ↗

Follow-up to Reviewer J1vB

Dear Reviewer J1vB,

We are writing to ask whether our response has addressed your concerns. We would be happy to discuss any remaining issues.

Best regards,

Authors of Submission #928

Reviewer_Gro4OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AHA, an automated red-teaming method for production coding agents such as Claude Code and Codex. Most prior methods keep the attacks that worked (prompts, seeds, strategy libraries). AHA instead keeps "vulnerability concepts": short explanations of why an attack works, when it should work, and what observation would prove the explanation wrong. These concepts are found by an autoresearch-style loop of LLM agents. One agent writes a hypothesis and its falsifier before any attack is designed, another builds the attack, a third checks the resulting trajectory against the falsifier, and a separate critic audits the process every 10 iterations. A concept is kept only after it has been confirmed several times. The evaluation covers 18 settings: three benchmarks (AgentHazard, AgentDyn, DTap), two agent harnesses, and three victim models. On held-out tasks, with a single attack per task, the kept concepts reach 47.0% attack success, compared with 32.8% for the best baseline, while using fewer victim queries. The paper also analyzes the types of vulnerabilities found, shows that concepts transfer to new models, harnesses and benchmarks, and shows that patches targeting the identified conditions reduce attack success on AgentHazard by about 41 points, more than generic guard models.

Critical strengths
  1. The idea is fairly novel. Storing the explanation of why an attack works, rather than the attack itself, is simple and useful. When a stored concept fails on a new task, you learn something (the condition was absent, or the explanation was wrong), and the enabling condition points directly at what to fix. The main design choices, such as writing the falsifier before the attack, requiring repeated confirmation, and the periodic critic, each address a real failure mode of long-running agent loops, like explaining results after the fact or gaming the judge.

  2. The comparison with AutoRISE* is convincing. It is the strongest baseline and runs on the same model as AHA, and AHA still wins clearly (47.0% vs. 32.8%). Concepts also hold up when moved to a different harness: concepts found on Claude Code reach 37.7% on Codex, better than any baseline searched directly on Codex.

  3. The evaluation is realistic and carefully engineered. Experiments run on real Claude Code and Codex agents in isolated containers, DTap uses live stateful backends, and held-out data is hidden from the discovery agents. The appendix includes the full prompts and end-to-end case studies, which helps reproducibility.

Critical weaknesses
  1. The novelty over closely related work is not clearly drawn, though I personally admit the novelty of this work. I understand that AutoDAN-Turbo already keeps a reusable library of natural-language attack strategies, and Claudini already uses autoresearch to discover new attacks. What seems genuinely new is the structured format (enabling condition, falsifier) and the repeated-confirmation gate. The paper should compare against this work explicitly, ideally with an ablation that swaps concepts for AutoDAN-Turbo-style strategy descriptions inside the same loop. I would be happy to see a more in-depth comparison between AHA and the related work.
  2. The main comparison is not fully fair. AHA's discovery loop runs on Claude-Opus-4.8 / GPT-5.5, while four of the five baselines (T-MAP*, IterInject*, AutoDAN-Turbo*, ReNeLLM*) generate their attacks with the weaker Qwen-3.7-Max. The paper does rerun AHA on Qwen (Table 2), but only for 3 of the 18 settings. In those settings AHA drops noticeably: on AgentHazard it falls from 85.6% to 70.0%, tying IterInject*, and on DTap its lead over the next-best baseline shrinks to a single instance. So part of the headline gain likely comes from the stronger model rather than from the concept representation itself.
  3. The baselines look weak. Simply replaying each benchmark's public attack payloads, with no search at all, averages 37.0% across the 18 settings, which is higher than every searched baseline (best 32.8%). On DTap the public payloads even slightly beat AHA itself under both harnesses. This suggests the adapted baseline implementations may be under-tuned, and the paper should discuss it.
Questions for authors
  1. Can you extend the same-backbone comparison (AHA@Qwen, or baselines on Opus-4.8 / GPT-5.5) beyond Deepseek/Claude Code, and report Q for AHA@Qwen?
  2. Why do the searched baselines underperform the benchmarks' public payloads on AgentHazard and DTap? Were baseline hyperparameters tuned per scenario?
4 comments
AuthorsOpenReview ↗

Response to Reviewer Gro4 (Part 1/3)

We sincerely thank the reviewer for the thoughtful and constructive review. We are encouraged that the reviewer finds the idea of storing the explanation of why an attack works simple and useful, and finds the comparison on real Claude Code and Codex agents convincing. We respond to the three concerns in order. The revised manuscript (PDF) is updated, with all changes highlighted in teal.


Q1: "The novelty over closely related work is not clearly drawn … AutoDAN-Turbo already keeps a reusable library of natural-language attack strategies, and Claudini already uses autoresearch to discover new attacks. … The paper should compare against this work explicitly, ideally with an ablation that swaps concepts for AutoDAN-Turbo-style strategy descriptions inside the same loop."

A1: Distinct from strategy libraries (AutoDAN-Turbo [2]) or automated search scripts (Claudini [3]), stored concepts retain falsifiable mechanisms rather than attack success alone. A concept commits its enabling condition and refuting observation before attack design; removing the falsifier lowers held-out ASR by 13.36 points. Gro4-Table 1 puts the three side by side (added to Appendix C.7 as Table 35).

Gro4-Table 1. AHA next to AutoDAN-Turbo and Claudini.

AutoDAN-Turbo [2]Claudini [3]AHA
What is kept for reuseA library of natural-language attack strategiesAttack algorithms as codeVulnerability concepts: mechanism, enabling condition, and falsifier
When an entry is addedWhen an attempt scores higher than the previous oneAt every iteration, each new variant is saved and ranked by its lossAfter the hypothesis and falsifier are written before the attack, and the concept is confirmed repeatedly
What a failed reuse tellsThe strategy did not workThe algorithm did not transferWhether the enabling condition was absent or the explanation was wrong
TargetChat LLMs, black-boxLLMs and safeguard models, white-boxProduction coding agents under Claude Code and Codex
Use for defenseNoneStress-tests proposed defensesEnabling conditions name the patch points

Swapping concepts for strategy descriptions inside the loop would change two things at once: what is stored, and how the loop runs, since the hypothesis and falsifier steps need the fields of a concept. We therefore test each change separately:

  • What is stored (new experiment). With the graph and the instantiator frozen, the instantiator receives the attack template alone. Held-out ASR drops by 9.12 to 11.65 points (added to Section 5.1).
  • How the loop runs. Removing the falsifier from the loop lowers held-out ASR by 13.36 points (Table 3).
  • Both. A strategy library in place of the whole loop is AutoDAN-Turbo* [2]. Under the same Qwen-3.7-Max researcher it reaches 33.60% against 42.39% for AHA.

Gro4-Table 2. Full concept and attack template alone. Held-out ASR (%) on Deepseek-V4-Pro.

HarnessScenarioFull concept ASR (%)Attack template alone ASR (%)
Claude CodeAgentHazard85.5670.00
AgentDyn37.5023.21
DTap59.1854.08
Mean60.7549.10
CodexAgentHazard91.1177.78
AgentDyn41.0725.00
DTap35.7137.76
Mean55.9646.84

Results analysis. Keeping only the attack template lowers held-out ASR from 60.75 to 49.10 under Claude Code and from 55.96 to 46.84 under Codex. Because this ablation removes only the mechanism and enabling condition inside the same frozen graph and instantiator, retaining falsifiable mechanism explanations adds transfer value beyond storing the attack itself. The gain appears in five of the six combinations, and DTap under Codex is level (35.71 vs. 37.76).

AuthorsOpenReview ↗

Response to Reviewer Gro4 (Part 2/3)

Q2: "The main comparison is not fully fair. AHA's discovery loop runs on Claude-Opus-4.8 / GPT-5.5, while four of the five baselines … generate their attacks with the weaker Qwen-3.7-Max … Can you extend the same-backbone comparison (AHA@Qwen, or baselines on Opus-4.8 / GPT-5.5) beyond Deepseek/Claude Code, and report Q for AHA@Qwen?"

A2: AHA's gains hold under an identical researcher backbone: we extend the experiment using Qwen-3.7-Max across all nine Claude Code settings, AHA reaches 42.39% mean ASR with 688 queries, outperforming searched baselines (added to Section 5.1 and Appendix C.2). At held-out time, one common host model (Claude-Opus-4.8 under Claude Code and GPT-5.5 under Codex) instantiates every stored artifact and runs AutoRISE* [1] (Appendix B.6). The other four baselines cannot run their attack generation on Claude-Opus-4.8 or GPT-5.5, so they generate with Qwen-3.7-Max, while AutoRISE* runs on the same host model as AHA (Appendix B.6).

Gro4-Table 3. Held-out ASR (%) under Claude Code on the three victims, with each method's researcher model, and total discovery queries Q over the nine settings. The rows marked Qwen-3.7-Max share one researcher model. AutoRISE* runs on Claude-Opus-4.8.

Method, researcher modelAgentHazard (3 victims, %)AgentDyn (3 victims, %)DTap (3 victims, %)Mean ASR (%)Total Q (queries)
T-MAP*, Qwen-3.7-Max55.56 / 46.67 / 65.5610.71 / 0.00 / 1.7939.80 / 22.45 / 50.0032.502,534
IterInject*, Qwen-3.7-Max55.56 / 50.00 / 70.003.57 / 3.57 / 5.3631.63 / 32.65 / 46.9433.251,820
AutoDAN-Turbo* [2], Qwen-3.7-Max54.44 / 45.56 / 57.785.36 / 1.79 / 23.2134.69 / 27.55 / 52.0433.603,365
ReNeLLM*, Qwen-3.7-Max46.67 / 42.22 / 55.561.79 / 0.00 / 0.0028.57 / 20.41 / 43.8826.572,074
AutoRISE* [1], Claude-Opus-4.858.89 / 48.89 / 62.227.14 / 0.00 / 26.7937.76 / 30.61 / 56.1236.49826
AHA, Qwen-3.7-Max66.67 / 58.89 / 70.0010.71 / 17.86 / 33.9339.80 / 30.61 / 53.0642.39688
AHA, Claude-Opus-4.875.56 / 71.11 / 85.5619.64 / 17.86 / 37.5046.94 / 36.73 / 59.1850.01594

Results analysis. With the same Qwen-3.7-Max researcher, AHA keeps an 8.79-point lead over the strongest Qwen-based baseline, and a 5.90-point lead over AutoRISE* running on Claude-Opus-4.8. It also uses fewer discovery queries than every baseline. The stronger researcher adds 7.62 points on top. The lead that remains under the same researcher comes from the concept-discovery loop, and it is largest on AgentHazard and AgentDyn.

AuthorsOpenReview ↗

Response to Reviewer Gro4 (Part 3/3)

Q3: "The baselines look weak. Simply replaying each benchmark's public attack payloads, with no search at all, averages 37.0% across the 18 settings, which is higher than every searched baseline (best 32.8%)." and "Why do the searched baselines underperform the benchmarks' public payloads on AgentHazard and DTap? Were baseline hyperparameters tuned per scenario?"

A3: Searched baselines face a generalization test, whereas Original replays instance-specific public payloads that already know each held-out target without discovery. Stored discovery artifacts receive only visible metadata under uniform, untuned hyperparameters:

Gro4-Table 4. Held-out ASR (%), model-averaged, for the public payloads, the strongest searched baseline in each combination, and AHA.

ScenarioHarnessOriginal ASR (%)Strongest searched baseline ASR (%)AHA ASR (%)
AgentHazardClaude Code65.1858.52, IterInject*77.41
AgentHazardCodex54.8264.45, T-MAP*74.81
AgentDynClaude Code10.7111.31, AutoRISE*25.00
AgentDynCodex5.366.55, IterInject*20.24
DTapClaude Code48.6441.50, AutoRISE*47.62
DTapCodex37.0734.69, AutoRISE*36.73

Why the comparison is set this way. The paper asks what a method retains from discovery, since a safety team reuses retained knowledge on new tasks after each deployment update. Every method therefore searches only on the discovery split and keeps one artifact. On each held-out instance it writes one attack from that artifact and the visible metadata, with no further search (Section 4.3). Original takes a different route: it replays the payload that the benchmark authors published for that exact held-out instance. It shows how hard each instance is, and it involves no transfer.

The baselines are not under-tuned. Each baseline follows the algorithm and scoring of its original paper and uses one configuration for all scenarios (Appendix B.6), under the same discovery splits, interfaces, judges, and per-attack limits as AHA. Each also uses more discovery queries than AHA over the nine Claude Code settings (826 to 3,365 against 594 to 688, Gro4-Table 3).

Why the public payloads score higher than the searched baselines. On DTap and on AgentHazard under Claude Code, Original is written for each held-out instance, while a searched method has to generalize from the discovery split.

AHA against the public payloads. AHA exceeds them by 12.23 to 19.99 points on AgentHazard and by 14.29 to 14.88 points on AgentDyn, and stays within 1.02 points of them on DTap. Averaged over the six combinations, AHA leads by 10.0 points (47.0% vs. 37.0%), so the headline gain does not rest on the strength of the searched baselines.


We hope these additions address the reviewer's main concerns, and we would be glad to provide further details.

References

[1] Gautam, Bahramali, and Atluri. AutoRISE: Agent-Driven Strategy Evolution for Red-Teaming Large Language Models. arXiv 2604.22871.

[2] Liu et al. AutoDAN-Turbo: A Lifelong Agent for Strategy Self-Exploration to Jailbreak LLMs. ICLR 2025.

[3] Panfilov et al. Claudini: Autoresearch Discovers State-of-the-Art Adversarial Attack Algorithms for LLMs. arXiv 2603.24511.

AuthorsOpenReview ↗

Follow-up to Reviewer Gro4

Dear Reviewer Gro4,

We are writing to ask whether our response has addressed your concerns. We would be happy to discuss any remaining issues.

Best regards,

Authors of Submission #928

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

For clarity and simplicity, we will refer to Reviewers Gro4, J1vB, and Js2s as R1, R2, and R3, respectively, in the following response.

We sincerely thank all reviewers for their thoughtful and constructive feedback. We are encouraged by their recognition of the key contributions and strengths of our work.

A novel and useful idea: "The idea is fairly novel. Storing the explanation of why an attack works, rather than the attack itself, is simple and useful" (R1). "The hypothesis-first design and repeated confirmation are particularly interesting" (R2). "The vulnerability concept can be useful" (R3).

A realistic and solid evaluation: "The evaluation is realistic and carefully engineered" (R1). "I especially appreciate that the authors go beyond ASR to study transferability and concept-guided defenses" (R2).

Core Contributions of Our Work

Vulnerability Concept As New Paradigm: We shift agent red-teaming from storing attack instances to discovering reusable, falsifiable vulnerability concepts that explain why an attack succeeds.

Hypothesis-Driven Autoresearch Loop: AHA writes each hypothesis and its falsifier before designing the attack, audits the logs with a fresh-context Critic, and keeps confirmed concepts in a Vulnerability Concept Graph.

Cumulative Transfer and Targeted Defense: Across 18 production settings, concepts raise held-out ASR by 14.2 points over the strongest searched baseline with fewer discovery queries, and enable a patch bundle that reduces AgentHazard ASR by 41.11 points.

We addressed every comment with new experiments or exact locations in the paper. Below we summarize the updates in our revision.


Updates of experimental results during Rebuttal

Main Text : Added controlled evaluations fixing the researcher model to Qwen-3.7-Max across all nine Claude Code settings (R1, R2, R3), statistical significance tests confirming AHA's gains over baselines in Section 5.1 (R3), an ablation comparing full concepts with attack templates alone (R1), and the cost of the patch bundle on benign user tasks (R3).

Appendix C.2 : Detailed same-backbone results with Qwen-3.7-Max and query counts over all nine settings (R1, R2, R3).

Appendix C.3 : Added attack template ablation on Codex and human audits on 100 reflections (R1, R2, R3).

Appendix C.6 : Added cross-family patch specificity experiments (Table 33) and benign task completion under the patch bundle (Table 34) (R2, R3).

Appendix C.7 : Added a side-by-side comparison with AutoDAN-Turbo and Claudini (Table 35) (R1).


Updates of in-depth discussions during Rebuttal

Main Text : Discussed baseline hyperparameter configurations and the role of public instance payloads in Section 4.3 (R1), clarified discovery query efficiency in Section 6 (R2), and renamed the ablation to "w/o confirmation gate" in Section 5.1 (R2).

Appendix C.2 : Discussed why public payloads outperform searched baselines on specific scenarios without tuning (R1), and reported the researcher sub-agent calls behind each discovery query (R2).

All revisions are highlighted in teal.


Summary of the Discussion: R2 confirmed that our response addressed their concerns and maintained the positive score. All three reviewers rate the paper positively, and our responses address the remaining questions of R1 and R3.

We sincerely thank the reviewers and the Area Chair for their careful handling of our submission, and we truly hope AHA can help make red-teaming of production agents a cumulative practice that explains why agents fail and where to repair them.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.