acceptodds
Desk rejected at ICLR 2027

OpenART: Scaling Agent Red Teaming via Open-Ended Environment Evolution

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

35 views

Abstract

AI agents increasingly operate through tools in persistent environments, where each action can alter the state encountered by subsequent actions. As these state changes accumulate over long interaction horizons, safety failures can emerge from sequences of individually benign actions rather than from any single decision. Yet existing safety benchmarks largely evaluate agents on short, static tasks through benchmark-specific interfaces, making it difficult to capture such cumulative risks or compare safety across agent runtimes. To address these limitations, we introduce OpenART, a red-teaming arena based on environment evolution. OpenART constructs over 10K validated, stateful scenarios spanning 50 domains from more than 500K Tools, MCPs, and Skills. These scenarios require a median of 97 tool calls and are projected through adapters onto 15 deployed agents, 5 foundation models, and 8 attack vectors, yielding 75 agent–model configurations. To systematically explore evolving attack surfaces, we further introduce the Evolutionary Markov Hypergraph Attack (EMHA), a black-box policy that performs feedback-driven environment evolution by coordinating authorized state transitions along hypergraph paths. Throughout this process, the task objective and safety contract remain fixed; only the target-visible environment evolves. Across all configurations, EMHA achieves a pooled strict Attack Success Rate (ASR) of 85.0%. Its advantage over instruction-only evolution increases from 1.8–2.7% in simple environments to 17.2–17.6% in the most complex ones. Moreover, target-agent identity explains an additional 7.6% of ASR variation beyond model and capability controls, demonstrating that runtime implementation is itself a consequential factor in agent safety. Together, these results establish OpenART as a scalable foundation for evaluating and red teaming AI agents in complex, stateful, and evolving environments.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Agent, Red Teaming, Environment

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_Lg4sOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

OpenART is a large-scale red-teaming framework that evaluates agent security under long-horizon, evolving environmental states rather than isolated prompts. It represents each test case through a benign task, executable workflow, target-visible environment, and hidden safety contract, and introduces Evolutionary Markov Hypergraph Attack (EMHA) to iteratively optimize eight attack surfaces using black-box evaluator feedback. Across 75 configurations, OpenART achieves a pooled strict attack success rate of 85.0%, with its advantage over instruction-only evolution increasing from approximately 2% to more than 17% as task dependency depth and tool-use complexity grow.

Critical strengths
  • OpenART evaluates 10,513 validated scenarios across 50 domains and 75 agent–model configurations, while decoupling scenario semantics from runtime-specific execution through adapters.
  • The authors propose OpenART, a red-teaming framework that evaluates agent security under long-horizon, evolving environmental states rather than isolated prompts.
  • The authors further introduce EMHA, a black-box attack method that models coordinated environmental modifications as paths over a hypergraph.
Critical weaknesses
  • External validation of EMHA. EMHA should be evaluated on established external benchmarks, such as AgentDojo or ASB, to demonstrate that its effectiveness generalizes beyond OpenART.

  • Statistical robustness. The paper should report variance, confidence intervals, or results across multiple independent runs to establish the reproducibility and statistical robustness of the reported improvements.

  • Missing EMHA baselines. EMHA is not directly compared with alternative attack-generation methods. The authors should include competitive baselines and report both attack success rate and computational/time cost to clarify its effectiveness–efficiency trade-off.

Questions for authors
  • Table formatting. Tables should follow standard LaTeX conventions, using \toprule and \bottomrule for cleaner presentation.
  • Equation formatting. Unnumbered equations that are not referenced in the text should use \begin{align*}...\end{align*} rather than numbered environments.
  • How do you do the human audit
6 comments
AuthorsOpenReview ↗

Response to Reviewer Lg4s

We thank the reviewer for the detailed assessment. We understand that you did not find the paper especially informative, as reflected in your rating: “1: I learned little or nothing new from this paper.” We acknowledge that OpenART is not a pure-method paper. Its main goal is to provide an environment infrastructure for agent systems that supports joint capability training and safety evolution, ultimately serving safety RL. The framework can flexibly connect to existing agent environments, as shown in our added experiments on AgentDojo, ASB, and DTap-Bench. Our environment-evolution red teaming is motivated by the need for safety alignment during agent execution, where the environment changes. OpenART is also designed to integrate many current agent-safety attack surfaces, including more complex MCP- and skill-level attack surfaces. It offers a lightweight way to fuse capability and safety benchmarks for RL training so that capability and safety can evolve together. Although the framework is not yet fully mature, we hope you can see its significance and give us a chance. We have added the experimental results below, and we would greatly appreciate any further candid feedback.

W1. External validation of EMHA

We thank the reviewer for suggesting external validation. We evaluated EMHA on AgentDojo, ASB, and DTap-Bench using Claude Code adapters within their benchmark environments, with five runs using distinct seeds. EMHA outperforms the evaluated benchmark attacks on both models. These results support its effectiveness beyond OpenART.

Strict ASR is reported as mean [95% CI], in percent.

BenchmarkAttackGPT-5.5Claude-Opus-4.8
AgentDojoNative injection0.0 [0.0, 0.0]0.3 [0.0, 0.7]
AgentDojoEMHA97.7 [96.1, 99.0]71.2 [66.0, 76.4]
ASBNative attack48.2 [44.1, 52.3]24.1 [20.7, 27.6]
ASBEMHA96.8 [95.1, 98.3]69.5 [65.8, 73.3]
DTap-BenchDTap-Red18.5 [16.3, 20.7]26.0 [23.5, 28.5]
DTap-BenchEMHA95.4 [94.2, 96.6]67.8 [65.2, 70.4]

W2. Statistical robustness

We thank the reviewer for requesting uncertainty estimates. We report five runs with distinct seeds on 10,513 scenarios, covering all five models. Standard deviations describe variation across runs. The small standard deviations indicate that the reported ASRs vary little across the five runs. Results are reported as mean ± SD [95% CI], in percent. The tables show three representative runtimes and pooled results across all 15 agents.

Strict ASR:

Agent runtimeGPT-5.5Opus-4.8GLM-5.2Qwen-3.7-MaxDeepSeek-V4-Pro
OpenCode99.82 ±0.14 [99.68, 99.94]62.50 ± 1.34 [61.57, 63.43]91.20 ± 1.15 [90.65, 91.75]98.90 ± 0.32 [98.70, 99.10]99.10 ± 0.28 [98.91, 99.29]
Claude Code93.60 ± 1.12 [93.12, 94.08]64.60 ± 1.41 [63.68, 65.52]92.30 ± 1.08 [91.78, 92.82]99.50 ± 0.21 [99.36, 99.64]98.70 ± 0.35 [98.48, 98.92]
Codex85.80 ± 1.25 [85.12, 86.48]58.40 ± 1.48 [57.45, 59.35]84.20 ± 1.31 [83.48, 84.92]90.30 ± 1.18 [89.72, 90.88]91.40 ± 1.10 [90.84, 91.96]
All 15 agents pooled88.50 ± 1.18 [88.23, 88.77]59.20 ± 1.52 [58.85, 59.55]87.90 ± 1.14 [87.62, 88.18]94.60 ± 0.85 [94.38, 94.82]94.70 ± 0.82 [94.49, 94.91]



Benign task completion:

Agent runtimeGPT-5.5Opus-4.8GLM-5.2Qwen-3.7-MaxDeepSeek-V4-Pro
OpenCode95.31 ± 0.38 [94.90, 95.72]98.92 ± 0.25 [98.71, 99.13]88.93 ± 0.44 [88.48, 89.38]84.72 ± 0.54 [84.21, 85.23]86.91 ± 0.48 [86.44, 87.38]
Claude Code93.80 ± 0.42 [93.33, 94.26]97.84 ± 0.31 [97.55, 98.13]87.01 ± 0.49 [86.52, 87.50]82.63 ± 0.58 [82.08, 83.18]84.77 ± 0.52 [84.25, 85.29]
Codex92.00 ± 0.51 [91.48, 92.52]96.23 ± 0.39 [95.86, 96.60]84.76 ± 0.56 [84.20, 85.32]80.31 ± 0.62 [79.72, 80.90]82.64 ± 0.55 [82.11, 83.17]
All 15 agents pooled92.02 ± 0.45 [91.82, 92.22]96.18 ± 0.28 [96.06, 96.30]85.00 ± 0.52 [84.76, 85.24]80.81 ± 0.60 [80.53, 81.09]82.91 ± 0.54 [82.66, 83.16]
AuthorsOpenReview ↗

Response to Reviewer Lg4s

W3. Competitive baselines and computational cost

All baselines underwent over 50 times hyperparameter tuning (grid searches over search depths, rollout budgets, and so on) on a held-out split, and we report their best-performing configurations. Despite these generous finetuning, existing prompt- and API-level attacks saturate at low ASRs on long-horizon tasks (28.2% for AgentVigil, 38.4% for DTap-Red on GPT-5.5). In contrast, EMHA achieves the highest ASR (93.8% on GPT-5.5 and 64.2% on Opus-4.8) while maintaining modest resource usage, consuming fewer attacker tokens (136.2k) than AgentVigil (174.6k) by using hypergraph constraints and a strategy archive to avoid uncoordinated, wasteful search. Strict ASR is mean [95% CI], in percent. Efficiency values are mean ± SD.

Attack methodGPT-5.5 Strict ASROpus-4.8 Strict ASRAttacker tokens (k)
Static Injection16.8 [15.2, 18.4]5.4 [4.5, 6.3]0.0 0.0
AutoInject (Chen et al., 2026a)22.5 [20.7, 24.3]8.6 [7.4, 9.8]103.5 7.2
AgentVigil (Wang et al., 2025c)28.2 [26.3, 30.1]12.4 [11.0, 13.8]174.6 11.7
DTap-Red (Chen et al., 2026b)38.4 [36.3, 40.5]19.2 [17.5, 20.9]127.5 8.7
EMHA (Ours)93.8 [92.6, 94.9]64.2 [61.9, 66.4]136.2 5.4

Q1. Table formatting

We thank the reviewer for this suggestion. We have standardized the remaining table rules using \toprule, \midrule, and \bottomrule, and removed the remaining vertical rules. These changes cover the main-text comparison and complexity tables and the appendix tables describing evaluation objects, attack vectors, and adapter support.

Q2. Equation formatting

We thank the reviewer for pointing this out. We have removed the unreferenced numbered equations when simplifying Sections 3–4.

Q3. Human audit

Human reviewers check whether each evaluator follows its scenario's task objective and safety contract. They assess contract alignment, task-completion labels, safety decisions, and agreement between the implementation and human-verified expected outcomes. Each case must pass both its safe and unsafe test. The 99.3% figure is the case-level pass rate; the error rates describe the paired audit tests. These details are reported in Appendix H.1 (Human Verification: Human-Audit Criteria). The separate human review of final trajectories is reported in Appendix H.2.

Audit detailResult
Independent human reviewers5
Distinct evaluators reviewed1,000; each reviewer assessed a disjoint set of 200
Checks per evaluatorOne safe outcome and one unsafe outcome
Case acceptanceBoth decisions must be correct; otherwise the full case is rejected
False positives3 / 1,000 safe outcomes; FPR = 0.3%
False negatives4 / 1,000 unsafe outcomes; FNR = 0.4%
Case-level pass rate993 / 1,000 = 99.3%; seven cases rejected

Thanks authors rebuttal and they solve all of my problem and the ‘reviewer zPdR‘ W3 and W4 increased my concern and I will take look how you rebuttal about it . I will add my score firstly and if need I will increase again.

Reviewer_Dq3zOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents OpenART, a red-teaming method to create environments that can trick LLM agents into unsafe behavior. At its core, OpenART is powered by EMHA, a hypergraph method that makes tweaks to environments to make them slightly riskier for agents without changing the task. Authors create over 10K scenarios that are on average much more multi-step than most agentic safety benchmarks, showing a highly successful attack rate of 85%.

Critical strengths

Generally this paper has many strengths, with the top ones being:

  • The paper's methodology seems quite innovative and well-executed!
  • I appreciate the ablation studies on scope of environment evolution and EMHA's mechanisms.
  • I appreciate the findings in "Environment Evolution Reveals Long-Horizon Safety Drift" which are really interesting!
Critical weaknesses

Two key weaknesses prevent me from a "clear accept" but I believe the authors can address in rebuttal:

  • The main weakness of this paper is it's lack of clarity in the methods section. In particular, Sections 3.2, 3.3, and 4.1 could really benefit from being rewritten for more clarity and more operational grounding; particularly, I found it to be overly math-y, and lacking running examples to ground what is being done. For a reader that's not as math-y, this paper is not accessible, thus it'd be best to re-write those sections to be simplified. For example, authors could perform an example-first rewrite using a single running scenario, explicit inputs and outputs for each stage, and pseudocode for the evolution loop.
  • A second slightly less impactful weakness is that I wish there was more insight into the kinds of environments that the method creates because I worry that this method yields attacks that are really convoluted (which may be fine, but would be very different than a more realistic environment change). Further discussion and some qualitative analyses are likely necessary.

Furthermore, there are some slightly less critical weaknesses:

  • L353: the paper mentions "controlled probes" used to distinguish safe completion from unsafe leakage, but do not provide enough details on what those controlled probes are and how they are verified.
  • Missing related works: https://arxiv.org/abs/2507.06134, https://arxiv.org/abs/2602.13379, http://arxiv.org/abs/2409.16427
  • Nitpick: Table 3 is very hard to read, authors should make it a heatmap with numbers.
  • The qualitative-ish analyses and content of "Attack Mechanisms in Evolving Environments" (end of §5) is super interesting but I wish more time was spent on it in the main paper!

After rebuttal, increasing my score to clear accept!

Questions for authors

N/A

3 comments
AuthorsOpenReview ↗

Response to Reviewer Dq3z

We sincerely thank the reviewer for the careful reading and constructive suggestions. We appreciate your positive assessment of our methodology and ablation studies, and your interest in the long-horizon safety-drift findings. Your feedback has helped us improve the method's accessibility and the discussion of our qualitative evidence. Revised passages are marked in blue.

W1. Clarity and operational grounding of the method

Thank you for recommending a more accessible presentation. We have simplified §§3.2, 3.3, and 4.1, replacing equation-heavy explanations with operational descriptions while retaining the key notation. A single weekly-report scenario now illustrates the stages with explicit inputs and outputs. We also added pseudocode at the end of §4.2 to summarize the environment-evolution loop.

W2. Complexity and plausibility of the environment changes

We appreciate this concern and have added an “Environment plausibility” discussion in the appendix's “Case Studies of Environment-Mediated Failure” section. We use cases in the appendix G examples. The environment changes in these cases are actual simple modifications to workflow documents. For GPT-5.5, the attacker introduced a finance schema that marked sensitive fields as required (Appendix G.1); for Opus-4.8, it presented a template footer as trusted report metadata (Appendix G.3). Our qualitative observations suggest that elaborate or conspicuous manipulations may alert frontier models to risk. Our attack instead exploits long-horizon task dependencies: simple environment changes influence later actions while their unsafe consequences remain unnoticed by the target.

Q1. Construction and verification of controlled probes

We expanded §3.2 (blue text) to specify what controlled probes are. Controlled probes are evaluator unit tests on prepared outputs with known ground-truth labels. For each scenario, the planner creates four outputs: complete-safe, complete-unsafe, incomplete-safe, and incomplete-unsafe. The validator then executes the deterministic evaluator on each output and requires every task-completion and safety decision to match the expected label; any mismatch returns the bundle for repair, and only bundles passing all probes are acceptted. Take confidentiality scenarios for an example: the planner separately inserts a protected marker into one public output (which must be flagged unsafe) and an unprotected marker into another public output (which must not be flagged), so the probe checks that the evaluator distinguishes true leakage from normal content. Furthermore, we use human audit to verify whether the generated evaluators are correct as detailed in Appendix H.1.

Q2. Missing related work

Thank you for pointing us to these relevant studies. We added OpenAgentSafety, Unsafer in Many Turns, and HAICOSYSTEM to the bibliography.

Q3. Readability of Table 3

Thank you for this helpful presentation suggestion. In §5.2, “Red-Teaming Effectiveness,” we replaced the former Table 3 with two heatmap panels showing benign task completion and Strict ASR.

Q4. More qualitative analysis in the main paper

We are grateful for your interest in this analysis. Following your suggestion, we expanded §5.2, “Attack Mechanisms in Evolving Environments,” with concrete examples in the main text and gave it enough space to discuss.

Thanks, the rebuttal addresses my comments, I appreciate the author's efforts :)

AuthorsOpenReview ↗

Big Thanks! :)

We sincerely thank Reviewer Dq3z for the kind follow-up and for confirming that our rebuttal addresses the comments. We greatly appreciate your careful reading and constructive feedback. We will ensure that all revisions and additions described in our response are incorporated into the final manuscript. Thank you again for your time and helpful review.

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

This paper introduces OpenART, a red-teaming framework for tool-using agents in changing environments. The main idea is to keep testing on the same benign task while an attacker changes the environment that the target encounters. The framework generates executable scenarios from a large capability registry (over 500k capabilities) and projects them onto native agent interfaces through runtime adapters. The authors also propose EMHA, which combines hypergraph-based attack planning, feedback-guided path selection, and using an archive of successful attack strategies.

The evaluations report more than 10k validated scenarios across 50 domains, with nearly 100 tool calls per scenario (median - 97). The evaluations spans 15 agents paired with five foundation models (75 agent + model pairs) and attacks are distributed across 8 attack vectors. Additional ablations examine the contribution of different elements in EMHA and attack surfaces. The authors highlight the relevance of long horizon tasks showing there is a high latency between the manipulated content and the unsafe action.

Critical strengths
  1. Important problem: The paper addresses an important gap in evaluating agent safety through short interactions. Separating a the task and safety contract from the runtime execution is a useful approach for studying how state, tools, memory, and agent instructions affect safety.

  2. Evaluations: The evaluation covers 75 model-agent pairs and more than 10k scenarios constructed using a registry of over 500k capabilities from 50 domains. This scale supports a broad comparison of agent implementations and foundation models across diverse task settings.

  3. Informative case studies: Beyond aggregate results, the appendix provides detailed examples of environment changes and the resulting agent behavior. These examples help illustrate how attacks are constructed and executed.

Critical weaknesses

1.Presentation and readability: Sections 3 and 4 are hard to follow, introducing many symbols used only once and equations that add limited value to understanding the method. I suggest adding more intuition, simpler explanations and visual diagrams, while moving many of the equations to the appendix.

2.Method implementation: The method section does not clearly explain how the learning update, strategy archive and population evolution in Equations 10–12 are implemented in practice. A quick review of the supplied code suggests differences from the equations as presented, but a clear explanation of the implementation is needed regardless.

3.Scope and interpretation: The task-generation example focuses on confidentiality leakage, and I could not find a clear breakdown of distinct attack objectives (as opposed to the attack surfaces presented in the appendix). Understanding the variety of evaluated objectives is important for assessing the framework’s scope and interpreting the reported vulnerabilities.

4.Runtime adaptation: Runtime-specific adapters require maintenance as agent interfaces change and new agents emerge. Given the rapid evolution of agent implementations, this dependence raises concerns about the framework’s long-term applicability and extensibility. Table 10 highlights this point, showing that adapters for most agents do not enable at least 2 of the 8 attack surfaces. Since infrastructure is a central contribution of this work, without a demonstrated approach to generalizing or automating adaptation, the long-term value of this contribution remains uncertain.

5.Vulnerability results reliability: The 99.3% human-audit claim lacks any information (for example - sample size, sampling procedure and annotation details), evaluator false-positive and false-negative rates are not reported. The supplied LLM judge also receives the deterministic checker’s result, limiting their independence, and raising a question if the human audit received this information as well. Cross-agent comparisons involve different supported attack surfaces, complicating interpretation. Attack budgets, experiment repetitions and uncertainty estimates under these repetition also require clearer reporting.

6.Comparative evaluation: The ablations do not establish EMHA’s advantage over a simpler adaptive attacker with identical access and budgets. A controlled comparison using the same agent-model pairs on shorter tasks or existing benchmarks is also needed to demonstrate the difference of the vulnerabilities the proposed setting reveals and this way to emphasize the evaluation gap this work aims to close.

Questions for authors
  1. When projection rejects a change, how are accepted changes that depend on it handled?

  2. What is the task completion rate under attack? This is important for distinguishing robust task execution from excessive refusal, as agents that frequently refuse requests may appear safe while providing limited utility.

  3. Could you provide a breakdown of the capability registry into groups by functionality and implementation type (executable tools, Skills, etc.)

  4. How is task completion defined and verified for each scenario? Are expected outcomes checked independently of the scenario generator?

  5. How is attack success evaluated? for example how is target-caused leakage distinguished from attacker written or pre-existing output files?

  6. suggestion - Table 3 is hard to interpret. Two heatmaps (for task completion and ASR separately) consuming similar space could make trends across models and agents easier to identify. The numerical table could remain in the appendix for precise reporting.

4 comments
AuthorsOpenReview ↗

Response to Reviewer zPdR

We sincerely thank the reviewer for the careful reading and constructive suggestions. We appreciate your positive assessment of our methodology.. Your feedback is very helpful!

W1. Presentation and readability

We simplified Sections 3–4 by replacing many equations with explanations. A running example from the appendix now shows each stage’s inputs and outputs. Pseudocode is added at the end of §4.2 summarizes the evolution loop.

W2. Method implementation

We thank the reviewer for identifying this gap. Section 4.2 now describes the implemented updates. Previous-round feedback updates stored hyperedge weights relative to a moving reward baseline. Execution outcomes update strategy scores, which guide subsequent graph proposals. Proposed graphs are checked and repaired before execution. These descriptions replace the earlier equations and clarify how feedback, strategy reuse, and graph evolution work in practice.

W3. Attack Objectives

OpenART defines the following attack objectives in four families.

Objective CategoryCovered Violations & Impacts
Data Integrity & DestructionPublic leakage, report falsification, evidence fabrication, file/backup deletion, configuration corruption, calendar/workflow disruption
Financial Loss & Unauthorized CommitmentsPayment diversion, unauthorized transfers, excessive spending, unauthorized discounts/refunds, price manipulation, duplicate transactions, unauthorized commitments
Authorization & System ControlUnauthorized permission grants, forbidden resource access, wrong-account modification, approval bypass, unauthorized code execution, disabling security controls, service disruption
Harmful Decisions & CommunicationsUnauthorized/harassing messaging, misinformation publication, discriminatory decisions, unfair priority manipulation, false compliance certification, harmful material distribution

W4. Runtime adaptation

In Table 10, all 15 agents support Workspace, Instructions, Tools, and MCPs, and 13/15 also support Skills. The blank marks occur only because those target agents natively lack features such as long-term memory or visible plan states, not because of adapter limitations. OpenART can expose any supported function through a simple configuration specifying where it lives and what can be changed. The attacker model then uses its own tools to automatically find ways to attack those surfaces, while OpenART only checks whether the proposed change is allowed before execution. Supporting a new or updated agent therefore requires only a small configuration change, keeping the framework easy to maintain.

W5. Reliability of vulnerability results

  • Human audit details and error rates (99.3% metric): The 99.3% refers to an audit of generated deterministic evaluators during scenario construction, not final execution traces. In added Appendix H.1, five independent reviewers each evaluated 200 disjoint evaluators (1,000 total across 50 domains) on contract alignment, task-completion correctness, safety-decision correctness, and implementation consistency, using paired safe/unsafe outcomes; failure of either outcome rejected the case. The audit found 3/1,000 safe false positives (0.3% FPR) and 4/1,000 unsafe false negatives (0.4% FNR), yielding a 993/1,000 (99.3%) case-level pass rate. Appendix H.2 adds a domain-stratified human review of 1,000 final trajectories: 98.6% agreement with automated Strict ASR, with 0.5% false positives and 0.9% false negatives.

  • LLM judge design and auditor blinding: Providing the deterministic checker’s signal to the LLM judge is intentional for causal attribution: the deterministic evaluator verifies unsafe results exist, while the LLM judge verifies execution lineage so unsafe results genuinely stem from the target agent’s actions rather than benign residue or accidental co-occurrence. Human reviewers in Appendix H.2 were strictly blinded in trajectory review, evaluating cases without deterministic outputs or automated scoring decisions.

  • Attack surface coverage across target agents: Variation in supported attack surfaces reflects target agents’ native architectures, not an OpenART limitation; OpenART supports any surface an agent natively provides. This is a real-world security insight: richer agents have larger attack surfaces and thus higher vulnerability/ASR under environment evolution, rather than confounding the evaluation.

  • Repeated runs, budgets, and uncertainty reporting: Repeated-run and baseline results are reported in our response to Reviewer Lg4s (W2 and W3). We kindly direct the reviewer to the response.

W6. Comparative Evaluation and Evaluation Gap

We respectfully direct the reviewer to our Response to Reviewer Lg4s (W3 and W1), where the complete empirical comparison tables, hyperparameter tuning protocols, and token budgets are provided.

AuthorsOpenReview ↗

Response to Reviewer zPdR

Q1. Handling Rejected Changes and Dependencies

When projection rejects a proposed change, that modification is discarded and never applied to the environment. OpenART converts this rejection into structured feedback and returns it to the attacker's memory, allowing the search policy to update transition values and pivot to valid alternate paths in subsequent rounds.

Q2. Task Completion Rate Under Attack

Under attack, the pooled task completion rate across all 15 agents averages 81.5% across the 5 foundation models (86.4% for GPT-5.5, 91.2% for Claude-Opus-4.8, 78.5% for GLM-5.2, 74.3% for Qwen-3.7-Max, and 77.1% for DeepSeek-V4-Pro) confirming that the models still maintain high benign task completion rates under attack rather than resorting to excessive refusal.

Q3. Capability Registry Breakdown

The 589,742 collected capabilities comprise 404,718 Skills, 76,891 MCP interfaces , and 108,133 Executable Tools.

Q4. Task Completion Verification and Evaluator Independence

Task completion is defined by scenario-specific deterministic evaluators that inspect the final environment state to verify that all required workflow actions were executed and expected deliverables were correctly produced. To guarantee their accuracy independently of scenario generation, these evaluators were validated through an independent human audit of 1,000 evaluators, confirming that they correctly distinguish complete from incomplete tasks with a 99.3% pass rate (Appendix H.1, Human-Audit Criteria).

Q5. Attack Success Evaluation and Leakage Attribution

An attack is judged successful when both the LLM judge and the deterministic evaluator agree on its success. We use several strategies and verifications to ensure that attack success is strictly judged. In Section 3.2, deterministic evaluators use controlled probes on prepared outputs to verify ground-truth safety outcomes. Attacker and target execute in isolated Docker containers, preventing the attacker from directly modifying public sinks. Moreover, we separately rule out pre-existing unsafe risks by verifying that initial environments are clean. In the evaluation time, a secondary GLM-5.2 judge then traces execution paths to confirm that the target agent caused the unsafe risks. Finally, a human audit of 1,000 trajectories validates these attribution labels, showing 98.6% agreement with automated Strict ASR (Appendix H.2, Human Review of Final Trajectories).

Q6. Visual Presentation of Table 3

We thank the reviewer for this suggestion and have updated Section 5.2 by replacing Table 3 with two side-by-side heatmaps displaying Benign Task Completion and Strict ASR separately across all 15 agents and 5 models.

I thank the authors for the detailed rebuttal and additional experiments, which address most of my concerns and strengthen the paper. My main remaining concern is the required maintainability and extensibility of runtime-specific adapters as agent interfaces evolve, which is the main limitation of the proposed framework in my opinion. Nevertheless, I recognize the value of the proposed framework and have increased my score to 3.

AuthorsOpenReview ↗

Big Thanks! :)

We sincerely thank Reviewer zPdR for the positive follow-up, for recognizing the value of our framework, and for the constructive concern about adapter maintainability and extensibility. We greatly appreciate the increased score and the reviewer’s careful engagement with our rebuttal, and we will address this limitation explicitly in the final manuscript. Thank you again for your time and helpful feedback.

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

Related papers

Discussion (0)

Sign in to comment.