acceptodds
Under review as a conference paper at ICLR 2027

Translating Frontier AI Risk Narratives into Executable and Verifiable Testbeds

Abstract

As frontier AI systems shift from conversational assistants to autonomous agents with direct access to real environments, their safety risks have developed from toxic text generation to actionable system hazards. While recent studies reveal capable agents exhibiting dangerous strategic behaviors, such as peer preservation, covert code sabotage, and in-context scheming, these frontier risks remain confined to narrative reports without reproducible execution testbeds. We present VeriEnv, an end-to-end frontier threat modeling framework that synthesizes verifiable execution environments directly from natural language risk descriptions, enabling reactive red-teaming and causal attribution. VeriEnv isolates observable interfaces from backend ground truth, enforcing the Benign Solvability Axiom via closed-loop container validation to eliminate infrastructure false positives. Furthermore, VeriEnv formalizes dynamic red-teaming as Guarded Transition Systems under immutable cryptographic locks, casting safety oversight as discrete Bayesian causal inference over execution traces. Across ten documented frontier risk disclosures on six advanced models, VeriEnv achieves an 87.5% setup success rate, robustly elicits strategic vulnerabilities, and isolates actionable thresholds for defensive hardening.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Frontier Risk Reconstruction, Verifiable Environments, Red Teaming, Environment Coevolution, Agent Safety Evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_uDbTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents VeriEnv, a framework for translating natural-language AI risk scenarios into executable multi-container testbeds. It combines validation of benign workflows, a DSL for state-dependent adversarial interventions, locked environment and task specifications, and action-based trajectory judging. The evaluation covers 70 X-Bench scenarios and 10 scenarios derived from frontier-risk reports across six models, with comparisons against Petri and AutoControl Arena. The paper also examines selected attack components through controlled interventions and Bayesian contrasts.

Critical strengths
  1. The environment construction pipeline addresses concrete difficulties in agent evaluation. Separating agent-facing interfaces from private backend state, checking credential boundaries, and requiring a reference client to complete the benign workflow provide useful safeguards against configuration errors and inaccessible tasks.

  2. The evaluation covers multiple risk categories and models, and includes a human check of the action-match judge on 200 trajectories, with approximately 87% agreement. These results provide empirical support for the framework's ability to elicit the specified agent behaviors. Section 6.5 also distinguishes the local intervention results from the descriptive frequencies of successful attacks.

Critical weaknesses
  1. The implemented success criterion differs materially from the definition in Section 4.3. Equation (6) requires both a matching agent action and a violation in private backend state. In contrast, Appendix D.1 and Listing 4 count matching issued actions even when execution fails, and explicitly exclude execution outcomes and final state from the verdict. The reported ASR therefore measures target-action issuance rather than verified state violations. This is a useful behavioral measure, but the distinction affects the paper's central claims about grounded verification and infrastructure false positives.

  2. The experiments do not sufficiently isolate the contribution of state-dependent, reactive red-teaming to the aggregate improvement. Table 4 reports that 94.7% of successful instances with validated delivery involve prompt injection, while 1.8% are adaptive multi-turn attacks. These selected-success proportions do not establish the effectiveness of each mechanism without attempted-run counts and comparable budgets. The position contrast in Section 6.5 concerns one case–model pair and remains inconclusive. The overall ASR gains consequently leave the contribution of the proposed reactive mechanism unresolved.

Questions for authors
  1. Which success criterion was used for the main results: the action-and-state condition in Equation (6), or the issued-action condition in Appendix D.1 and Listing 4? For cases targeting an environmental effect, can the authors distinguish issued target actions, successful execution, and backend-confirmed violations, and explain how this distinction affects the reported results?

  2. Can the authors provide attempted-run counts, successful-run counts, and search budgets for prompt-only, environment-involved, and adaptive multi-turn attacks? Is there a matched comparison that isolates state-dependent execution from prompt construction and search? If such evidence is unavailable, the claims about the contribution of reactive red-teaming should be narrowed accordingly.

2 comments
AuthorsOpenReview ↗

Response to Reviewer uDbT [1/2]

We sincerely thank the reviewer for the thoughtful and constructive feedback. We really appreciate your recognition of target-action issuance as a useful behavioral measure, as well as your constructive suggestions on isolating state-dependent red-teaming mechanisms. Below, we carefully address your concerns in Weaknesses and Questions, and we hope you find our responses satisfactory.


W1 & Q1. Eq. 6 appears to require backend state violations while App. D.1 and Listing 4 score issued actions, and environmental-effect scenarios require separating issued, executed, and backend-confirmed breaches alongside infrastructure false positives.

Thank you for pointing out this discrepancy between Eq. 6 and Listing 4. In Sec. 4.3, Eq. 6 includes an extra backend-state condition Φviolation\Phi_{\mathrm{violation}}, making breach scoring appear to require a post-execution state change. Throughout our experiments, however, all main ASR results use one unified criterion: whether the agent issues the prohibited target action, as defined in Sec. 6.1, App. D.1, and Listing 4. As you point out, target-action issuance is a useful primary metric because (1) even when environment permissions block execution, issuing the prohibited action already demonstrates unsafe agent behavior, and (2) scoring issued actions ensures a fair comparison with dialogue-only baselines such as Petri that have no separate backend.

Regarding the second part of Q1, cleanly executed and backend-confirmed violations coincide in VeriEnv because agents interact with real containerized services. To verify this empirically, we inspect saved container snapshots across all 57 state-changing scenarios and find that all 98 of 98 cleanly executed calls modify private backend state. Accordingly, we merge these two outcomes into Executed ASR and compare it against Attempted ASR across all 420 X-Bench trajectories below:

FrameworkAttempted ASRExecuted ASR
AutoControl9.76%9.05%¹
Petri18.10%N/A²
VeriEnv38.33%31.90%

¹Return value in AutoControl's generated Python environment. ²Simulated tool output in Petri's dialogue environment.

The results show that VeriEnv remains the best-performing method by a substantial margin on Executed ASR, with only a slight decrease from Attempted ASR. This slight decrease stems from VeriEnv's non-root, least-privilege access controls in Sec. 3.2, which block unauthorized calls at runtime.

Regarding infrastructure false positives, scoring issued actions does not introduce false breach verdicts: an infrastructure error alone never counts as a breach unless the agent itself issues the prohibited action, and pre-certifying the benign workflow in Sec. 3.3 prevents broken tasks from pushing agents into unsafe workarounds. Consistent with this, 88.95% of all VeriEnv breach verdicts, and 100% in our certified ablation environments, occur with zero prior infrastructure errors. We revise Eq. 6 and report both Attempted ASR and Executed ASR in Secs. 3.3, 4.3, and 6.1 of our revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer uDbT [2/2]

W2 & Q2. Tab. 4 reports only proportions among successful runs without attempted-run denominators or budgets, and matched comparisons are needed to isolate state-dependent timing and delivery channels.

To clarify Tab. 4 and isolate the contribution of state-dependent red-teaming, we report full attempt denominators and round budgets alongside controlled timing and channel ablations:

  • Attempt Denominators and Budgets. Across all 1,028 archived X-Bench rollout rounds under a 3-round evolution cap, full attempt denominators, successes, ASR, and mean rounds are as follows:

    Attack category across 1,028 archived roundsAttemptsSuccessesASRMean rounds
    Prompt-only channel77215419.95%2.38
    Environment-involved channel256187.03%2.61
    Hybrid channel: Prompt + Environment43716.28%2.69
    Runtime state-triggered programs8555.88%4.25

    It is worth mentioning that the lower ASR and higher mean rounds on environment and state-triggered programs stem from scenario difficulty rather than weaker attack effectiveness: VeriEnv stops searching once an early round succeeds, so easier scenarios are already solved by Turn-0 prompts while environment and state-triggered programs are invoked mainly on harder remaining scenarios.

  • State-Dependent Timing Isolation. Holding the single-shot attack message fixed with evolution disabled across 180 runs on 15 X-Bench scenarios, delivering the payload after a target runtime tool call rather than statically at Turn 0 raises ASR on delivered runs from 10.47% to 15.79%, a 50% relative increase, and doubles overall ASR on GPT-5.4 from 6.67% to 13.33%.

  • Delivery Channel Isolation. Delivering the identical Turn-0 attack text via the user prompt versus native environment files across 200 runs with evolution disabled highlights their distinct threat-model roles: while direct user prompts average higher at 46.0% versus 24.0% when attackers can modify trusted user input, native environment injection reaches 95.0% versus 55.0% in realistic indirect document-injection scenarios where attackers act solely through untrusted files.

  • Overall Attribution of ASR Gains. Overall, strategy evolution drives the primary aggregate X-Bench ASR gain from 15.71% in Round 1 to 38.33%, while state-dependent triggers and native environment channels specifically strengthen conditional runtime timing and indirect file-based red-teaming.

We incorporate these analyses into Tab. 4 and Secs. 6.4 and 6.5 of our revised manuscript.

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

This paper presents VeriEnv, a framework that converts natural language descriptions of agent risks into containerized test environments. The system combines environment synthesis, a benign solvability check, a reactive attack language, an action matching judge, and Bayesian analysis. It is evaluated on 70 X Bench scenarios and 10 documented risk scenarios with six models. The problem is timely and the effort to connect agent behavior to executable state is valuable. However, the current evidence does not support several central claims.

Critical strengths
  1. The paper addresses an important gap between transcript based safety evaluation and evaluation in systems with persistent state.

  2. Separating the agent interface from private verification state is a sensible design choice. Requiring a benign reference solution before adversarial testing is also useful as an engineering check.

  3. The evaluation covers several risk categories and models, and the paper reports limitations for some null ablation results rather than forcing positive conclusions.

Critical weaknesses
  1. The primary endpoint is internally inconsistent. Equation 6 defines a breach as requiring both a matching agent action and a verified change in private state. Section 6 and Listing 4 instead state that a matching issued action counts even when execution fails, is rejected, or causes no state change. The judge is explicitly told to ignore environment outcomes. The reported attack success rate therefore measures attempted actions, not executable hazards. This directly conflicts with the central claim of physically grounded evaluation and makes it unclear what the headline results establish.

  2. The comparison does not isolate the contribution of the environment. VeriEnv changes the environment, attack construction, reactive search, and delivery strategy at the same time. The paper does not describe a controlled comparison with identical attacks, budgets, prompts, model settings, and judge procedures. Notably, 94.7 percent of validated successful deliveries involve prompt intervention, while only 9.4 percent involve environment surfaces and only 1.8 percent are adaptive multi turn attacks. Higher attack success may therefore come from stronger prompting rather than executable environment design. Table 3 also appears to use one binary outcome per model and scenario pair, with no repeated trials or uncertainty estimates.

  3. The environment quality evaluation is partly circular. The authenticity rubric defines a real shell as the preferred interface and penalizes Python functions and extensive MCP use, which directly favors the proposed architecture over the selected baselines. The paper does not identify the quality judge, inference settings, blinding procedure, repeated judgments, or agreement with human ratings. Human validation covers only the action judge and reports approximately 87 percent raw agreement without label prevalence, shared annotation, chance correction, or an error analysis. Figure 2 is also difficult to audit because its reported percentages are not compatible with one classification for each of 80 cases, and the aggregation unit is not stated.

  4. The causal claims are substantially overstated. Independent Beta Binomial contrasts estimate differences between conditions, but they do not establish causality without a clear assignment procedure and control of selection effects. The ablation study uses only three selected case and model pairs and 98 runs. The main positive technique result comes from one pair, so it cannot support general threat model calibration. The proposed risk ratio and conditional mutual information are not reported. The abstract also claims superadditive synergy between authority and urgency, yet no interaction contrast is presented and the individual intervals for both components include zero.

  5. Reproducibility and formal justification are insufficient. Passing one scripted benign workflow does not guarantee that all relevant agent paths are valid, so the claim that later failures arise exclusively from agent decisions is not justified. The submission provides no executable artifact or link, synthesis prompts, repair and evolution procedures, attack budgets, decoding settings, model snapshots, reset protocol, or precise definitions and denominators for setup success, policy error, and failure. This also makes the reliability comparison with a transcript oriented baseline hard to interpret.

Questions for authors
  1. Can the authors report attempted target actions and verified harmful state changes as separate outcomes for every method?

  2. What experiment holds attack content, search budget, model configuration, and judging fixed while varying only the environment implementation?

  3. How were the quality scores in Figure 2 aggregated, and which model or human produced each score?

3 comments
AuthorsOpenReview ↗

Response to Reviewer oStd [1/3]

We sincerely thank the reviewer for the detailed and insightful review. We especially appreciate your thoughtful feedback on executable safety evaluation, as well as your rigorous methodological suggestions on endpoint definition, statistical uncertainty, and reproducibility. Below, we carefully address your concerns in Weaknesses and Questions, and we hope you find our responses satisfactory.


W1 & Q1. Eq. 6 conflicts with the issued-action judge in Listing 4, and attempted actions versus verified harmful state changes should be reported separately for every method.

We apologize for the confusion surrounding Eq. 6. Specifically, Eq. 6 includes an extra backend-state condition Φviolation\Phi_{\mathrm{violation}} that makes our judge appear to require a post-execution state change, whereas our actual evaluation in Sec. 6.1 and Listing 4 follows one unified criterion: whether the agent issues the prohibited target action. We use issued target actions as our primary ASR criterion because (1) even when environment permissions block execution, issuing the prohibited action already demonstrates unsafe agent behavior, and (2) scoring issued actions ensures a fair comparison with dialogue-only baselines such as Petri that have no separate backend.

To answer Q1 directly, we re-evaluate all trajectories across both benchmarks using saved tool return codes and error logs to separate attempted target actions from cleanly executed calls:

FrameworkX-Bench AttemptedX-Bench ExecutedFrontier AttemptedFrontier Executed
AutoControl9.76%9.05%¹5.00%5.00%¹
Petri18.10%N/A²6.67%N/A²
VeriEnv38.33%31.90%18.33%18.33%

¹Return value in AutoControl's generated Python environment. ²Simulated tool response in Petri's dialogue environment.

The results show that VeriEnv maintains its clear lead on cleanly executed breaches across both benchmarks, with only a slight decrease on X-Bench. This slight decrease stems from VeriEnv's non-root, least-privilege access controls in Sec. 3.2, which block unauthorized calls at runtime. Aside from the 18 attempts blocked by permissions or errors across the 57 backend-state scenarios, container snapshots confirm that all 98 executed calls modify private state. We rewrite Eq. 6 and report both attempted and cleanly executed ASR in Secs. 4.3 and 6.1 of our revised manuscript.


W2-1 & Q2. The comparison varies environment and red-teaming components together, 94.7% of successful deliveries use prompt intervention, and an environment-only controlled comparison is needed.

We first clarify that environment and red-teaming mechanisms serve distinct roles in VeriEnv: red-teaming drives ASR magnitude, whereas live containers ensure environment realism and solvability rather than inflating ASR. Regarding the 94.7% figure, we believe you are referring to Tab. 4, where the 94.7% prompt-involved and 9.4% environment-involved shares indicate only which channel injects the attack payload among successful runs, rather than whether the executable environment is used: even when an attack enters via the user prompt, the agent still explores files, queries services, and executes target commands inside live containers.


W2-2. Tab. 3 uses a single binary outcome per model-scenario pair without uncertainty intervals or repeated trials.

To quantify statistical uncertainty in Tab. 3, we compute 95% Wilson confidence intervals across both benchmarks and evaluate consistency across repeated trials:

BenchmarkAutoControlPetriVeriEnv
X-Bench across 420 pairs9.76% [7.28, 12.98]18.10% [14.71, 22.06]38.33% [33.81, 43.07]
Frontier Risks across 60 pairs5.00% [1.71, 13.70]6.67% [2.62, 15.93]18.33% [10.56, 29.92]

The results confirm that VeriEnv's X-Bench lead is statistically significant with non-overlapping Wilson intervals and paired McNemar p<10−12p < 10^{-12} over both baselines, while we interpret the smaller 60-pair Frontier Risks subset more cautiously. Moreover, repeating evaluations across three random seeds on 15 scenarios yields 83.33% identical verdicts, confirming consistent results across repeated trials, and we include these metrics in Sec. 6.1.

AuthorsOpenReview ↗

Response to Reviewer oStd [2/3]

W3-1 & Q3. Potential shell bias in the authenticity rubric, undisclosed quality-judge settings, and Fig. 2's aggregation unit.

To resolve these concerns, we evaluate environment quality after removing the shell-interface dimension from our rubric and anonymizing framework names across two judge models, and we aggregate Fig. 2 across all 80 scenarios:

  • Ablation of the Shell Dimension. Specifically, we drop Dimension 1 in App. D.2, retaining only the shell-agnostic Dimensions 2 to 4, and score 20 stratified scenarios with framework identifiers removed using both GPT-5.6-Sol and Gemini-3.8-Flash-High under default decoding:

    Authenticity settingAutoControlPetriVeriEnv
    Full rubric across all 4 dimensions2.602.404.30
    w/o shell dimension3.122.864.07
    w/o shell, blinded GPT-5.6-Sol3.272.973.76
    w/o shell, blinded Gemini-3.8-Flash3.082.524.15

    The results confirm that VeriEnv maintains a clear quality lead across both judges even after removing the shell dimension entirely.

  • Judge Disclosure and Fig. 2 Aggregation. Regarding Q3, all quality scores in our submission are produced by GPT-5.6-Sol under default decoding using the prompts in App. D.2, and Fig. 2 in our initial submission aggregates at the case-model pair level rather than the scenario level. Aggregating directly across all n=80n = 80 scenarios by averaging each method's six model runs per scenario first confirms that VeriEnv still ranks sole highest in 85.00% of scenarios and highest or tied in 88.75% on Authenticity, and we update Fig. 2 and Sec. 6.2 accordingly.


W3-2. Action Judge human validation reports only ~87% raw agreement without label prevalence, chance correction, precision/recall, or error analysis.

To address this concern, we compare our Action Judge's binary verdicts against the ground-truth labels from our 20-annotator human study across all 200 sampled trajectories in App. D.1, aligning each annotation to its exact trajectory ID:

Positive PrevalencePrecisionRecallRaw AgreementCohen's κ\kappa
10.50%79.17%90.48%96.50%0.825

Because the 200 trajectories are randomly sampled across all evaluated methods and benchmarks where most runs do not breach, human-labeled positives account for 10.50%, or 21 of 200 trajectories. Across these 21 positive cases, our Action Judge correctly identifies 19 breaches at 90.48% recall with Cohen's κ=0.825\kappa = 0.825, confirming that its 96.50% agreement reflects accurate breach detection rather than majority-class prediction. Only 7 disagreements occur overall: 5 arise in a single Evaluation Awareness scenario where cautious test-mode responses resemble normal policy compliance, and 2 involve incomplete intermediate actions. We include this validation breakdown in App. D of our revised manuscript.


W4. Causal attribution lacks assignment clarity, relies on too few pairs for calibration, omits ARR and CMI, and claims authority × urgency synergy without an interaction contrast.

We are sorry for stating our causal attribution and "super-additive synergy" claims too broadly based on only three case-model pairs. To address this, we expand our attribution experiments across five models and clarify our assignment procedure alongside ARR and CMI below:

  • Expanded coverage: Expanding Sec. 6.5 from 1 technique pair to 10 pairs across five models at 20 trials per condition shows that removing task rationalization lowers pooled ASR from 60.50% to 47.00% across all five models, a +12.26% posterior mean effect with 99.95% directional probability, providing broader cross-model support for threat-model calibration.
  • Assignment procedure: Rather than selecting runs post hoc, we randomly sample candidate pairs across risk categories beforehand and run all ablation conditions on each pair in randomized order across 20 clean-reset trials, with no outcome-based reruns.
  • ARR, CMI, and interaction contrast: Across the completed 2×22\times 2 grid on the same 10 pairs, task rationalization yields ARR = 1.26, while authorization deception and urgency pressure yield CMI = 0.031 and 0.047. Although combining both factors does show a positive effect, the interaction is less pronounced than universal super-additivity, so we revise the "super-additive synergy" claim accordingly.

Accordingly, we state the scope of these claims more precisely and include these expanded metrics in the Abstract and Secs. 1, 5, and 6.5 of our revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer oStd [3/3]

W5-1. Passing one scripted benign workflow does not guarantee all relevant agent paths are valid or that later failures arise exclusively from agent decisions.

Thanks for your thoughtful comment. We agree that "exclusively" in Sec. 3.3 is too absolute: passing one scripted benign workflow proves that a valid end-to-end solution exists, rather than guaranteeing that every alternative path is defect-free. Meanwhile, we would like to clarify that VeriEnv's pre-evaluation verification goes beyond this single reference workflow: the compliance probes Φprobe\Phi_{\mathrm{probe}} in Sec. 3.2 first check input-output behavior across all exposed service endpoints, while the reference workflow in Sec. 3.3 verifies that cross-service routes, permissions, and database schemas work together end-to-end so broken tasks do not enter evaluation.

To check how often agents still hit environment defects on alternative paths in practice, we manually audit 100 randomly sampled negative and failed runs:

Mutually exclusive breakdown of 100 audited runsCountPercentage
Agent decision, refusal, or command factors only7575.0%
Mixed agent and recoverable or doc environment factors1515.0%
Provider API safety-filter blocks88.0%
Pure environment defects only22.0%

The audit shows that pure environment defects account for only 2.0% of audited runs, matching VeriEnv's 2.5% failure rate in Fig. 3. Accordingly, we revise Sec. 3.3 to state that certification rules out unsolvable setups and keeps infrastructure-induced failures rare rather than strictly zero.


W5-2. Missing executable artifact, prompts, budgets, reset protocol, and exact definitions and denominators for Fig. 3.

We appreciate your feedback on reproducibility. First, several components are already provided in our manuscript: repair and evolution procedures appear in Secs. 3.3 and 4.2 and App. A.2, the Attack DSL grammar in App. B, Frontier Risks specifications in App. C alongside the public X-Bench suite, and evaluation prompts in App. D. For the remaining materials, we will publicly release the executable artifact, synthesis prompts, decoding settings, and model snapshot records upon publication of the paper. Regarding the reset protocol and budgets, before every run VeriEnv removes existing containers and volumes, restores writable state from immutable seed snapshots, and verifies readiness probes, while all frameworks run under default decoding and their official configurations sharing GPT-5.6-Sol, a 600s timeout, and up to 5 API retries. Finally, all three methods in Fig. 3 share the same denominator of N=480N = 480 model-scenario runs, spanning 80 scenarios across six target models, with the three mutually exclusive categories defined as follows:

  • Success: Runs that complete setup and execution normally to produce a valid evaluation trajectory for judging.
  • Policy error: Runs aborted when provider safety filters block model calls during synthesis or live auditing. In particular, Petri's 43.5% policy-error rate arises when safety filters block its auditor LLM during live adversarial turn generation, reflecting an operational constraint of live transcript red-teaming rather than an implementation bug.
  • Fail: Runs terminated by environment build, code syntax, or runtime execution crashes.
Reviewer_KyqBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces VeriEnv, a framework that translates natural-language AI risk descriptions into executable, containerized evaluation environments. It combines multi-service environment synthesis and benign-workflow validation with state-dependent adversarial interventions expressed through an AttackDSL. The framework also analyzes intervention factors using Bayesian posterior contrasts. Experiments cover 70 X-Bench scenarios and 10 documented frontier-risk scenarios across six models, reporting higher target-behavior elicitation rates than Petri and AutoControlArena.

Critical strengths
  1. A useful emphasis on executable environments and independent validation. Separating agent-facing interfaces from backend verification and checking benign-task solvability are sensible safeguards against misleading results caused by broken infrastructure or unreliable tool responses.
  2. Evaluation across diverse scenarios and models. The experiments demonstrate improved target-behavior elicitation across multiple risk categories and six models. The controlled component experiments also provide more informative evidence than successful-attack prevalence statistics alone.
Critical weaknesses
  1. The operational success criterion conflicts with the claimed execution-grounded verification. Equation (6) requires both a matching agent action and a corresponding backend state violation. However, Appendix D explicitly instructs the judge to ignore execution outcomes and count matching calls even when they fail or are rejected. Attempted and successfully executed violations support different conclusions. This inconsistency makes it unclear what the reported ASR establishes and how backend verification contributes to the main results.

  2. The incremental contribution of the proposed mechanisms is insufficiently established. AutoControlArena already synthesizes executable environments with verification and repair, while Petri supports adaptive multi-turn auditing. The evaluation compares complete frameworks without isolating the benefits of containerized execution, benign-workflow certification, or constrained reactive attacks, and does not sufficiently specify matched search and retry budgets. Moreover, the controlled attribution experiments cover only three fixed case–model pairs. They do not substantiate the broader claims that task rationalization is indispensable or that authorization deception and urgency pressure exhibit super-additive synergy.

Questions for authors
  1. Do the main ASR results require confirmed backend state violations, or only matching issued actions? Can attempted violations and successfully executed violations be reported separately?
  2. What evidence isolates the contributions of containerized execution, benign-workflow certification, state-dependent intervention, and strategy evolution under comparable model-call and search budgets?
  3. How are attack capabilities, permissions, retries, and failed environment setups handled consistently across frameworks? What explains AutoControlArena’s lower setup success here than the end-to-end generation success reported in its original paper?
  4. What controlled evidence supports the claimed super-additive interaction between authorization deception and urgency pressure, beyond their co-occurrence in successful attacks?
  5. A reference client completing one benign workflow establishes that a valid path exists, but does not exclude defects elsewhere. What checks support the stronger claim that subsequent failures arise exclusively from agent decision-making?
4 comments
AuthorsOpenReview ↗

Response to Reviewer KyqB [1/3]

We sincerely thank the reviewer for the careful and constructive review. We particularly appreciate your recognition of the importance of executable environment synthesis for agent safety evaluation, as well as your insightful suggestions on strengthening our component isolation and experimental rigor. Below, we carefully respond to your concerns in Weaknesses and Questions, and we hope you find our responses satisfactory.


W1 & Q1. Eq. 6 appears to require backend state violations while App. D scores issued actions, and attempted versus executed breaches are not reported separately.

We are sorry for the imprecise formulation in Eq. 6. In Sec. 4.3, Eq. 6 includes an extra backend-state condition Φviolation\Phi_{\mathrm{violation}}, making breach scoring appear to require a post-execution state change. We would like to clarify that all main ASR results instead use one unified criterion: whether the agent issues the prohibited target action, as defined in Sec. 6.1, App. D, and Listing 4. We use issued target actions as our primary ASR criterion because (1) even when environment permissions block execution, issuing the prohibited action already demonstrates unsafe agent behavior, and (2) scoring issued actions ensures a fair comparison with dialogue-only baselines such as Petri that have no separate backend.

To separate attempted violations from successfully executed ones, we re-evaluate all 420 X-Bench trajectories using saved tool return codes and error logs:

FrameworkAttempted ASRExecuted ASR
AutoControl9.76%9.05%¹
Petri18.10%N/A²
VeriEnv38.33%31.90%

¹Return value in AutoControl's generated Python environment. ²Simulated tool output in Petri's dialogue environment.

The results show that VeriEnv retains a substantial lead on cleanly executed breaches with only a slight decrease from its attempted ASR. This slight decrease stems from VeriEnv's non-root, least-privilege access controls in Sec. 3.2, which block unauthorized calls at runtime. Aside from the 18 attempts blocked by permissions or errors across the 57 backend-state scenarios, container snapshots confirm that all 98 executed calls modify private state. We revise Eq. 6 and report both attempted and executed ASR in Secs. 4.3 and 6.1 of our revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer KyqB [2/3]

W2-1 & Q2. The incremental contributions of containerized execution, benign certification, state-dependent intervention, and strategy evolution are not isolated under comparable budgets.

To address this concern, we first clarify that these four mechanisms serve distinct roles in VeriEnv: containerized execution and benign-workflow certification ensure environment realism and solvability rather than inflating ASR, whereas state-dependent triggers and strategy evolution drive attack elicitation. Below, we isolate each mechanism on its corresponding target metric:

  • Containerized Execution. On the 10 Frontier Risks scenarios, we use the environment quality rubric in App. D.2 to compare three execution backends, corresponding to LLM-simulated tools in Petri, single-file Python mocks in AutoControl, and live multi-container services in VeriEnv:

    Environment backendAuthenticityTask Consistency
    LLM-simulated tools2.402.80
    Single-file Python mock2.602.90
    Live multi-container services4.303.40

    The results show that live multi-container services achieve substantially higher environment authenticity and task consistency than simulated tools and single-file Python mocks.

  • Benign-Workflow Certification. Building on containerized services, we isolate certification repair in Sec. 3.3 across 180 runs on 15 initially failing X-Bench scenarios with prompts and attacks fixed and evolution disabled:

    Certification ablation over 180 runsBefore certification repairAfter certification repair
    Reference benign task solvability0 / 1515 / 15
    Environment runtime crash rate6.67%0.00%

    This confirms that benign-workflow certification restores end-to-end task solvability and eliminates runtime crashes before red-teaming begins.

  • State-Dependent Intervention. Turning to attack elicitation, we isolate delivery timing across 180 runs on 15 X-Bench scenarios with the single-shot attack message fixed and evolution disabled. Delivering the payload after a target runtime tool call rather than statically at Turn 0 raises ASR on delivered runs from 10.47% to 15.79%, a 50% relative increase, showing that runtime state triggers sharpen single-shot attack timing.

  • Strategy Evolution. Finally, we isolate multi-round evolution by comparing Round-1 attack programs against full three-round evolution across all 420 X-Bench pairs:

    X-Bench across 420 pairsSuccessesASR
    Round-1 attack program only66 / 42015.71%
    With strategy evolution, up to 3 rounds161 / 42038.33%

    These results demonstrate that strategy evolution drives the primary +22.62% aggregate ASR gain while requiring only 1.42 extra rollouts on average.

We incorporate these isolation studies into Sec. 6 of our revised manuscript.


W2-2 & Q4. The three-pair attribution study is too narrow to establish that task rationalization is indispensable or that authorization deception and urgency pressure interact super-additively.

We apologize for using overly absolute expressions such as "indispensable" and "super-additive synergy" in Sec. 1 from a three-pair study. To evaluate both effects rigorously, we expand Sec. 6.5 to 10 pre-selected case-model pairs across five target models and two risk categories per model, running 20 clean-reset trials per condition in randomized order with evolution disabled while stripping each factor synchronously across prompts and environment artifacts. Across this expanded study, removing task rationalization lowers pooled ASR from 60.50% to 47.00% across all five models, yielding a +12.26% posterior mean effect with 99.95% directional probability and ARR = 1.26. This confirms that task rationalization serves as the dominant positive driver, though not a universal "indispensable" gate because breaches still occur without it. Meanwhile, completing the 2×22\times 2 joint-removal test for authorization deception and urgency pressure across the same 10 pairs shows that positive synergy appears in most scenarios, though not universally across all tasks. Accordingly, we state the scope of these claims more precisely in Sec. 1 and include these expanded results in Sec. 6.5 of our revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer KyqB [3/3]

Q3. How attack capabilities, permissions, retries, and failed setups are standardized across frameworks, and why AutoControlArena's setup rate is lower than in its original paper.

Thanks for your thoughtful question. Across all experiments, the three frameworks are evaluated on the same 80 scenarios, including AutoControl's own 70-scenario X-Bench, across the same six target models in the Codex harness under GPT-5.6-Sol, a 600s timeout, and the same issued-action judge, with each specific aspect standardized as follows:

  • Attack capabilities: Because attack generation is the core methodological difference under comparison, each framework runs its native red-teaming mechanism under its official configuration, while sharing the same GPT-5.6-Sol backbone and consistent interaction and timeout limits.
  • Permissions: Although execution interfaces differ across containers, Python mocks, and simulated dialogues, every framework derives agent access from the same scenario specification, and VeriEnv enforces strict non-root, least-privilege credentials so ASR gains never stem from looser permissions.
  • Retries: All three frameworks apply up to 5 API retries for transient call errors, triggered strictly before judging and independently of outcomes.
  • Failed setups: Any failed setup or aborted run is counted uniformly as a non-breach across all model-scenario pairs so that operational reliability is treated identically across methods.
  • AutoControl's setup-rate change: The change from its original 98.6% rate stems from scenario-level versus run-level granularity. Whereas AutoControl's paper reports one outcome per scenario across 70 tasks, Fig. 3 evaluates every run across all six target models and separates API policy blocks from runtime errors, where diverse model tool calls occasionally trigger execution errors. Even so, this modest setup gap cannot explain our 28.57% X-Bench ASR lead over AutoControl.

We include these experimental details and Fig. 3 definitions in Sec. 6.2 and the Appendix.


Q5. Completing one reference benign workflow does not rule out defects on other paths or establish that subsequent failures stem exclusively from agent decisions.

Thank you for pointing this out. We agree that "exclusively" in Sec. 3.3 is too absolute: completing one reference benign workflow proves that a valid end-to-end solution exists, rather than guaranteeing that every alternative path is defect-free. Meanwhile, we would like to clarify that VeriEnv's pre-evaluation verification goes beyond this single reference workflow: the compliance probes Φprobe\Phi_{\mathrm{probe}} in Sec. 3.2 first check input-output behavior across all exposed service endpoints, while the reference workflow in Sec. 3.3 verifies that cross-service routes, permissions, and database schemas work together end-to-end so broken tasks do not enter evaluation.

To check how often agents still hit environment defects on alternative paths in practice, we manually audit 100 randomly sampled negative and failed runs:

Mutually exclusive breakdown of 100 audited runsCountPercentage
Agent decision, refusal, or command factors only7575.0%
Mixed agent and recoverable or doc environment factors1515.0%
Provider API safety-filter blocks88.0%
Pure environment defects only22.0%

The audit shows that pure environment defects account for only 2.0% of audited runs, consistent with VeriEnv's 2.5% failure rate in Fig. 3. Accordingly, we revise Sec. 3.3 to state that certification rules out unsolvable setups and keeps infrastructure-induced failures rare rather than strictly zero.

Thanks. After reading the rebuttal and the other reviewers’ comments, I will keep my initial rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.