acceptodds
Under review as a conference paper at ICLR 2027

HazardAuditor: From Executable Threats to Safer Computer-Use Agents

Abstract

Computer-use agents increasingly interact with browsers, terminals, file systems, and external services, introducing safety risks that emerge through runtime behavior rather than generated content alone. Existing guard models target static prompts and responses and are poorly suited to agent execution; existing executable safety platforms produce evaluation verdicts rather than the normalized supervision a guard model needs to learn across heterogeneous agent frameworks. We introduce HAZARDAUDITOR, an execution-grounded framework that closes both gaps. Its infrastructure runs heterogeneous agents (Claude Code, Codex, Hermes, and OpenClaw) in controlled environments and normalizes their interactions into a canonical event representation for cross-framework supervision. We further observe that token-level post-training objectives create a structural mismatch for generative guards, causing longer rationales to dominate gradient updates. Guard Policy Optimization (GuardPO) addresses this by converting deterministic safety outcomes into sequence-level advantages and normalizing rationale and verdict regions, making the safety decision the effective unit of optimization. Across multiple benchmarks and heterogeneous computer-use systems, HAZARDAUDITOR improves accuracy by up to 16.5 percentage points over the strongest prior guard. Code, models, and evaluation artifacts will be available at https://anonymous.4open.science/r/HazardAuditor.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agent safety, agent guard model, computer-use agents

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_rtKTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces HazardAuditor, an execution-grounded safety framework that normalizes diverse execution logs from heterogeneous computer-use agents into a unified behavioral representation to detect runtime risks. To prevent variable-length explanations from dominating model training, the authors propose Guard Policy Optimization (GuardPO), which normalizes the rationale and verdict regions to directly optimize the sequence-level safety decision.

Critical strengths
  • Comprehensive experiments with consistent performance improvement: The paper evaluates HazardAuditor across five distinct benchmarks (CUA-EXEC, AgentHazard, ATBench, R-Judge, and ASSE-Safety) and four heterogeneous agent frameworks (Claude Code, Codex, Hermes, and OpenClaw). Across these diverse settings, the proposed framework demonstrates robust and consistent gains, notably improving accuracy by up to 16.5 percentage points over the strongest prior guard model on the execution-grounded diagnostic.

  • Sufficient baseline comparison with the latest model variants: The authors establish a rigorous evaluation setting by comparing against a comprehensive suite of various types of models. This includes new decision models such as Jev and general-purpose safety guards (e.g., Llama Guard 3, Qwen3Guard), and recent trajectory-aware agent guardrails (e.g., AgentDoG, ShieldAgent, and BraveGuard),

Critical weaknesses
  • Incremental Novelty of Guard Policy Optimization (GuardPO): While GuardPO effectively mitigates the rationale length-bias issue in generative guards, its core algorithmic novelty is highly incremental compared to established policy optimization formulations. The method fundamentally relies on the existing Clipped Importance Sampling Policy Optimization (CISPO) framework to handle token-level policy clipping and sequence-level advantage broadcasting. The primary structural modification is restricted to applying a separate length-normalization across the rationale and verdict regions before loss aggregation. From a policy gradient derivation perspective, this acts as a domain-specific masking and weighted averaging heuristic for the loss implementation rather than a fundamental theoretical advancement in reinforcement learning.

  • Simplistic Reward Formulation and Advantage Allocation: The sequence-level advantage calculation relies on a highly simplistic, deterministic reward structure (+1 for correct, -1 for incorrect, -1.25 for malformed) based purely on the terminal verdict. Because this single sequence advantage is uniformly broadcast to every token in the response, the optimization process does not explicitly evaluate or shape the quality, logical soundness, or factual grounding of the intermediate reasoning steps. Exploring relative advantage calculations such as comparing multiple reasoning trajectories against one another for the same prompt or incorporating dense rewards would have provided a much more robust and rigorous validation of the proposed policy optimization approach.

Questions for authors

Do you have any ablation on the reward design?

6 comments
AuthorsOpenReview ↗

Rebuttal 1

We sincerely thank the reviewer for the thoughtful assessment and for recognizing the execution-grounded motivation, comprehensive evaluation, and decision-level optimization of HazardAuditor. We appreciate the insightful comments on the novelty of GuardPO and its reward formulation. We provide the following clarifications and additional experimental evidence.

Q1: GuardPO appears algorithmically incremental over CISPO, with its primary modification being region-wise loss normalization. How does it establish a distinct contribution?

A1: We agree that GuardPO adopts the clipped importance-weighting principle of CISPO for token-level policy updates. Our contribution is a decision-aligned optimization objective addressing a structural mismatch in generative safety guards. Each response represents one safety decision, whereas conventional token-level aggregation allows longer rationales to exert disproportionate influence.

GuardPO addresses this through sequence-level outcome advantages and separate response-level normalization of rationale and verdict losses. To isolate these contributions, we compare standard GRPO, which uses group-relative advantages and conventional token-level aggregation, against GuardPO and its normalization ablations under the same SFT initialization and training budget.

Table 1: GuardPO ablation on CUA-EXEC.

MethodRationale Norm.Verdict Norm.Accuracy (%)
HazardAuditor-SFTNoNo80.50
Standard GRPONoNo81.27
GuardPO w/o Rationale Norm.NoYes87.62
GuardPO w/o Verdict Norm.YesNo81.33
GuardPO (Full)YesYes90.88

Standard GRPO improves over SFT by only 0.77 percentage points, whereas GuardPO improves over GRPO by 9.61 points. Removing rationale normalization reduces accuracy by 3.26 points, while removing verdict normalization leads to a 9.55-point reduction. These results demonstrate that the gains are not simply attributable to reinforcement learning, but arise substantially from decision-aligned loss aggregation.

Q2: The reward formulation is simplistic, and the sequence-level advantage is uniformly broadcast to all response tokens. Do you have any ablation on the reward design? Why not use dense rewards or relative advantages?

A2: Our reward design follows the operational objective of a safety guard. Unlike an agent policy that optimizes sequential environment interactions, HazardAuditor performs a classification task and must produce a correct binary safety verdict. We therefore use deterministic terminal rewards rather than introducing additional learned reward models or LLM judges.

Independently verifiable process-level reward for individual reasoning steps are not available. Dense rewards would require such supervision and could incentivize plausible-sounding explanations without improving final decision correctness. We instead use SFT to establish evidence-grounded reasoning behavior and GuardPO to directly optimize the terminal safety decision.

We further examine reward sensitivity by comparing alternative deterministic reward assignments under the same optimization objective.

Table 2: Reward-design ablation on CUA-EXEC.

Reward VariantCorrectIncorrectMalformedAccuracy (%)
Binary outcome reward+1-1-189.91
Malformed-neutral reward+1-1084.26
Stronger format penalty+1-1-1.590.35
GuardPO (Ours)+1-1-1.2590.88

Removing the malformed-output penalty reduces accuracy by 6.62 points. Treating malformed outputs as ordinary errors achieves 89.91%, while increasing the penalty to -1.5 achieves 90.35%. Our original setting performs best among the evaluated variants, suggesting that a moderate format penalty provides a more effective balance.

Regarding relative advantages, Table 1 already provides a direct comparison with standard GRPO, which computes group-relative advantages across multiple rollouts for the same prompt. Its substantially lower accuracy further supports our decision-level optimization design. Uniformly broadcasting the sequence advantage is deliberate because the reward evaluates one complete safety decision, not independently verifiable reasoning tokens. Region-level normalization prevents rationale length from determining each response's optimization weight.

Q3: Beyond GuardPO, what is the substantive contribution of the execution infrastructure?

A3: HazardAuditor combines two complementary contributions: execution-grounded supervision and decision-aligned optimization. The infrastructure goes beyond format conversion by establishing an adaptive red-teaming and executable verification pipeline.

Safety-aligned agents frequently refuse static harmful requests. Our Red Agent therefore follows an observe–adapt–act loop, reformulating requests, decomposing objectives, escalating gradually, and changing injection points in response to refusals. We evaluate benign operation, direct attacks through user messages, and indirect attacks additionally injected through tool observations.

Table 3: Adaptive red-teaming coverage and attack success rates.

Evaluation ConditionAttack SurfaceVerified ASR
Benign operationNo adversarial interventionN/A
Direct attackAdaptive user-message injection90.6%
Direct + indirect attackUser messages and injected tool observations93.9%
Increment from indirect surfaceAdditional compromised observations+3.3 pp

Representative attack strategies achieve ASRs of 88.5% for task decomposition, 87.9% for obfuscation, 85.5% for instruction injection, and 83.7% for jailbreaks. ASR measures attack success against target agents, not HazardAuditor's detection performance.

Importantly, attack generation and outcome verification are separated. Executable verifiers inspect MCP records and sandbox state rather than relying on LLM judgments alone. They reject 7.30% of Control-LLM-positive outcomes, and 99.88% of verifier programs reference execution logs or environment-state evidence. Behavioral labels additionally capture unsafe attempts even when the intended environmental effect is not achieved, consistent with our original labeling protocol.

The infrastructure covers 126 risk categories, 77 attack techniques, 30 environment categories, and 71 implemented tool APIs across four heterogeneous agent frameworks. Thus, its contribution lies in systematically generating adaptive attacks and producing independently verifiable, execution-grounded behavioral supervision, rather than merely unifying existing logs.

We sincerely thank the reviewer again for the constructive feedback. We hope the additional ablations clarify the specific contributions of GuardPO, while the infrastructure analysis demonstrates the broader methodological contribution of HazardAuditor. We would be happy to address any remaining questions and provide further clarification. If these concerns have been adequately resolved, we respectfully hope the reviewer might reconsider the overall assessment of our submission.

We greatly appreciate your time and constructive feedback. Should you have any questions regarding our rebuttals, please feel free to contact us, we would be more than happy to discuss them with you and make further revisions as needed.

Reviewer_gioTOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper adapts trajectories from four agent frameworks into a unified event representation with behavior-level safety labels, then trains Qwen3Guard-Gen-8B with SFT followed by GuardPO, a sequence-level outcome reward with region-normalized loss. It reports clear gains on its own CUA-Exec benchmark, but only parity with BraveGuard on AgentHazard and a slight regression on R-Judge.

Critical strengths

The paper builds guard training at the decision level so that one response corresponds to exactly one safety decision, and this training approach is both novel and effective. The authors also release the trained model together with the training and evaluation code, which makes the method reproducible and directly usable.

Critical weaknesses
  1. The data-side contribution is limited. Converting, unifying, and labeling trajectories from several common agent frameworks is engineering integration rather than a new supervision mechanism.
  2. GuardPO shows clear improvements, but the validation is insufficient. There is no comparison against alternative update rules and no component ablation, and the gains largely disappear on external benchmarks, with one backend even regressing, so generalization is not established.
Questions for authors

Could the authors provide an ablation comparing GuardPO against other update rules under the same data and compute budget, to establish the necessity of the decision-level update?

7 comments
AuthorsOpenReview ↗

Rebuttal to #gioT#

We sincerely thank the reviewer for recognizing the novelty and effectiveness of our decision-level guard training approach, as well as the importance of releasing reproducible artifacts. We appreciate the concerns regarding the data-side contribution and the validation of GuardPO. We provide additional evidence and clarifications below.

Q1: The data-side contribution appears to be primarily engineering integration rather than a new supervision mechanism.

A1: We agree that schema conversion alone constitutes engineering integration. However, our contribution goes substantially beyond format unification. The infrastructure establishes an adaptive, multi-surface, and verifier-grounded supervision pipeline.

First, safety-aligned target agents frequently refuse static harmful requests. Our adaptive Red Agent therefore follows an observe–adapt–act loop, reformulating requests, decomposing objectives, escalating gradually, and changing injection points in response to refusals. This generates meaningful execution trajectories rather than merely collecting refusal responses.

Second, we evaluate three distinct conditions covering different trust boundaries: benign operation, direct attacks through user messages, and indirect attacks additionally injected through tool observations.

Table 1: Adaptive red-teaming coverage and attack success rates.

Evaluation conditionAttack surfaceVerified ASR
Benign operationNo adversarial interventionNormal control
Direct attackAdaptive user-message injection90.6%
Direct + indirect attackUser messages and injected tool observations93.9%
Additional indirect attack surfaceIncrement over direct attacks+3.3 pp

Representative adaptive attack strategies achieve ASRs of 88.5% for task decomposition, 87.9% for obfuscation, 85.5% for instruction injection, and 83.7% for jailbreaks.

Third, our labeling pipeline does not rely on an LLM judge as the final authority. The Control LLM proposes potential attack successes, but every positive judgment is subsequently checked by an executable verifier against MCP records and sandbox state. The verifier validates environmental outcomes, whereas behavioral safety labels are independently assigned based on the agent's attempted behavior, including unsafe attempts that do not achieve their intended environmental effects. Specifically, the verifier rejects 7.30% of LLM-proposed successes.

The infrastructure further covers 126 risk categories, 77 attack techniques, 30 environment categories, and 71 implemented tool APIs across four heterogeneous agent frameworks. Thus, the contribution is not simply integrating existing logs, but systematically generating adaptive attacks, collecting execution-level evidence, and producing independently verifiable behavioral supervision.

Q2: Could the authors provide an ablation comparing GuardPO against alternative update rules under the same data and compute budget?

A2: Thank you for this suggestion. We conduct additional controlled experiments comparing standard GRPO with GuardPO and its two normalization ablations. All variants share the same SFT initialization, training data, rollout budget, and evaluation protocol. Standard GRPO uses conventional group-relative advantages and token-level loss aggregation. All RL variants share the same SFT initialization, training data and training compute budget.

Table 2: Ablation of GuardPO on CUA-EXEC.

MethodRationale NormalizationVerdict NormalizationAccuracy (%)
HazardAuditor-SFTNoNo80.50
Standard GRPONoNo81.27
GuardPO w/o Rationale NormalizationNoYes87.62
GuardPO w/o Verdict NormalizationYesNo81.33
GuardPO (Full)YesYes90.88

Standard GRPO improves over SFT by only 0.77 percentage points, whereas GuardPO improves over GRPO by 9.61 points. Removing rationale normalization reduces accuracy by 3.26 points, while removing verdict normalization results in a larger decrease of 9.55 points. These results establish that the gains are not simply attributable to reinforcement learning, but arise substantially from aligning optimization with response-level safety decisions.


We sincerely thank the reviewer again for the constructive feedback. We hope the additional evidence clarifies that our data-side contribution extends beyond engineering integration and that the controlled ablations establish the independent value of GuardPO. We believe these clarifications and results substantially address the concerns regarding the novelty of the supervision pipeline and the validity of the proposed optimization method.

We would be happy to address any remaining questions. If these concerns have been adequately resolved, we respectfully hope the reviewer will reconsider the overall assessment of our submission.

We greatly appreciate your time and constructive feedback. Should you have any questions regarding our rebuttals, please feel free to contact us, we would be more than happy to discuss them with you and make further revisions as needed.

Reviewer_gioTOpenReview ↗

Response to rebuttal

Thank you for the additional ablation. Two concerns remain.

1.Generalization: the new ablation is reported only on CUA-Exec. On the external benchmarks in the paper, GuardPO's gains over SFT are much smaller than on CUA-Exec and sometimes negative, comparable to the gain of standard GRPO there. Could you report the GRPO vs. GuardPO comparison on these external benchmarks, and confirm that HazardAuditor is scored on exactly the same test trajectories, with the same metric definitions, as the reported baselines?

2.Novelty: the adaptive red-teaming, executable verifiers, and ASR results described in A1 closely match those in Vera, and BraveGuard already trains trajectory-level guards on executed computer-use rollouts. Could you clarify which components of your pipeline are new relative to these two works?

I am willing to raise my score if these concerns are addressed.

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

This paper presents HazardAuditor, an execution-grounded safety framework for computer-use agents. It collects and normalizes execution traces across four agent frameworks and proposes Guard Policy Optimization (GuardPO) to optimize safety decisions while reducing the influence of rationale length. Experiments demonstrate promising improvements across multiple agent safety benchmarks.

Critical strengths
  1. Clear and well-grounded motivation. The paper highlights an important distinction between merely encountering harmful content and actually attempting unsafe actions, showing why agent safety assessment requires execution-level evidence rather than content-based classification alone.

  2. Unified cross-agent supervision. The proposed infrastructure normalizes execution traces from four heterogeneous agent frameworks, enabling a single guard model to learn from different execution environments.

  3. Comprehensive experimental evaluation. The paper evaluates HazardAuditor on both the proposed CUA-Exec benchmark and four existing agent safety benchmarks, with extensive analyses covering cross-framework performance, precision-recall trade-offs, etc.

Critical weaknesses
  1. Insufficient ablation of GuardPO. Although GuardPO substantially improves performance over the SFT baseline, the experiments do not clearly isolate the benefits of its proposed response-level normalization from those of outcome-based reinforcement learning in general.

  2. Limited validation of runtime safety. Although the paper positions HazardAuditor as a runtime guard, the experiments primarily evaluate safety classification on recorded execution trajectories. It remains unclear whether the model can identify unsafe behavior early enough to prevent harmful actions from being executed.

Questions for authors
  1. How does GuardPO compare with standard policy optimization methods without the proposed response-level normalization on rational and verdict?

  2. Can HazardAuditor reliably detect unsafe behavior before harmful tool actions are executed? Could the authors provide experimental results evaluating its early detection capabilities?

  3. How robust is HazardAuditor against adversarial attacks, such as prompt injections embedded in tool outputs or deliberately obfuscated unsafe actions?

5 comments
AuthorsOpenReview ↗

Rebuttal to #QGng# part1

We sincerely thank the reviewer for the positive assessment of our work, particularly for recognizing its execution-grounded motivation, unified cross-agent supervision, and comprehensive experimental evaluation. We also appreciate the constructive questions regarding the methodological contribution of GuardPO, runtime safety intervention, and robustness against adversarial attacks. We have conducted additional experiments and analyses to address these concerns.

Q1: How does GuardPO compare with standard policy optimization methods without the proposed response-level normalization of rationale and verdict tokens?

A1: Thank you for identifying this important gap. We agree that the original comparison between SFT and GuardPO does not fully isolate the contribution of response-level normalization from the general benefits of outcome-based reinforcement learning.

We therefore conduct additional ablations on CUA-EXEC, comparing standard GRPO with GuardPO and its normalization variants. All methods share the same SFT initialization, training data, rollout budget, and evaluation protocol. Standard GRPO uses group-relative advantages and conventional token-level loss aggregation, without the separate response-level normalization introduced by GuardPO.

Table 1: Ablation of GuardPO on CUA-EXEC.

MethodRationale NormalizationVerdict NormalizationAccuracy (%)
HazardAuditor-SFTNoNo80.50
Standard GRPONoNo81.27
GuardPO w/o Rationale NormalizationNoYes87.62
GuardPO w/o Verdict NormalizationYesNo81.33
GuardPO (Full)YesYes90.88

The results reveal three important findings. First, standard GRPO improves accuracy by only 0.77 percentage points over SFT, whereas full GuardPO achieves a 10.38-point improvement. Second, removing rationale normalization reduces accuracy from 90.88% to 87.62%, indicating its contribution to preventing variable-length explanations from dominating optimization. Third, removing verdict normalization reduces accuracy to 81.33%, highlighting the importance of explicitly normalizing the final safety decision.

These results support our central hypothesis that aligning optimization with response-level safety decisions provides benefits beyond standard outcome-based reinforcement learning. In particular, both normalization components contribute to the observed improvement, with verdict normalization having a larger impact.

Q2: Can HazardAuditor reliably detect unsafe behavior before harmful tool actions are executed? Could the authors provide experimental results evaluating its early detection capabilities?

A2: We appreciate this important concern. We agree that post-hoc trajectory classification alone is insufficient to establish the practical value of a runtime guard. To address this, we conduct an additional pre-execution evaluation on A3S-Bench.

Unlike conventional trajectory-level evaluation, our evaluation assesses the guard at actionable decision points during agent execution. The model receives the interaction context and proposed tool invocation, including its arguments, and must identify unsafe behavior before the corresponding tool action is dispatched. Future tool outputs and subsequent execution events are excluded to prevent information leakage.

Table 2: Early Detection and Pre-execution Interception on A3S-Bench.

ModelFirst-Exposure Warning (%) ↑Eventual Detection (%) ↑Final Macro-F1 (%) ↑Benign False Block (%) ↓
HazardAuditor (Ours)48.067.064.927.0
AgentDoG-Qwen3-4B38.058.054.596.0
Llama-Guard-3-8B1.04.035.51.0
Qwen3Guard-Gen-8B1.02.035.03.0

HazardAuditor achieves the highest first-exposure warning rate (48.0%), eventual detection rate (67.0%), and final Macro-F1 (64.9%) among the evaluated models. Notably, it substantially outperforms AgentDoG in early warning and eventual detection while maintaining a considerably lower benign false-block rate (27.0% versus 96.0%). In contrast, the general-purpose guards exhibit very low false-block rates but detect very few unsafe interactions.

These findings demonstrate a more favorable balance between early threat detection and benign-action preservation. Although early detection remains challenging, the results provide direct evidence that HazardAuditor can identify unsafe behavior before harmful tool actions are executed, rather than serving exclusively as a post-hoc trajectory auditor. Actual blocking is handled by an external execution controller using the guard's verdict.

Together with our existing NSFA query-level evaluation, these experiments provide complementary evidence of risk detection at both the request level and the tool-execution level.

AuthorsOpenReview ↗

Rebuttal to #QGng# part2

Q3: How robust is HazardAuditor against adversarial attacks, such as prompt injections embedded in tool outputs or deliberately obfuscated unsafe actions?

A3: Thank you for raising this important concern. Our framework is specifically designed to distinguish malicious content encountered during execution from unsafe behavior initiated by the agent. This distinction is particularly important for prompt-injection attacks, where malicious instructions may appear in otherwise legitimate tool outputs.

To clarify the breadth of our execution infrastructure, we provide additional statistics on its risk taxonomy, attack techniques, environment coverage, and tool surface.

Table 3: Coverage and Scale of the HazardAuditor Infrastructure.

DimensionCurrent ScaleOperational Evidence
Risk taxonomy126 risk categoriesAll canonical categories represented in infrastructure-derived training data
Attack taxonomy77 attack techniquesAll canonical techniques represented in infrastructure-derived training data
Environment coverage30 environment categoriesCoverage across all 30 semantic environment categories
Containerized execution15 container definitions; 12 active per run10 backend services, 1 MCP gateway, and 1 agent container
Tool surface71 implemented tool APIsTool interactions normalized into the canonical event representation

The infrastructure supports four heterogeneous agent frameworks, namely Claude Code, Codex, Hermes, and OpenClaw. Its coverage spans diverse risk categories, attack techniques, and execution environments, providing behaviorally grounded supervision beyond isolated harmful prompts.

Furthermore, the A3S-Bench evaluation in Table 2 provides an additional adversarial assessment under injection-based attacks. HazardAuditor achieves a 48.0% first-exposure warning rate and a 67.0% eventual detection rate, outperforming the compared guards. Importantly, our labeling protocol distinguishes between an agent that merely encounters malicious instructions and one that follows them to initiate unsafe actions.

Reproducibility and Open-Source Commitment

We are committed to making this work fully reproducible. We will publicly release the trained model weights, complete training and optimization framework, source code, evaluation scripts, configurations, and relevant evaluation artifacts. This release will enable the community to reproduce our results, inspect the implementation of GuardPO, and extend the framework to additional agent architectures and execution environments.

We sincerely thank the reviewer again for the constructive feedback and positive assessment. The additional GRPO ablations clarify the contribution of GuardPO's response-level normalization, while the A3S-Bench evaluation provides direct evidence of HazardAuditor's pre-execution detection capabilities. The expanded infrastructure analysis further clarifies the diversity of risks and attack techniques represented in our execution-grounded supervision.

We believe these additions substantially address the concerns regarding methodological attribution, runtime applicability, and adversarial evaluation. We would be happy to provide further clarification if any questions remain. If the additional evidence adequately resolves the concerns raised, we respectfully hope the reviewer might reconsider the overall assessment of our submission.

We greatly appreciate your time and constructive feedback. Should you have any questions regarding our rebuttals, please feel free to contact us, we would be more than happy to discuss them with you and make further revisions as needed.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal

Dear Reviewers,

We sincerely thank all reviewers for their constructive feedback and recognition of HazardAuditor's contributions. We conducted additional experiments and analyses to address the concerns raised during the discussion.


Reviewer QGng

  • Added controlled GRPO comparisons and GuardPO ablations. GuardPO achieves 90.88% accuracy, compared with 81.27% for standard GRPO and 80.50% for SFT.
  • Added pre-execution evaluation on A3S-Bench. HazardAuditor achieves 48.0% first-exposure warning, 67.0% eventual detection, and 64.9% Macro-F1.
  • Clarified runtime intervention and expanded adversarial coverage analysis across 126 risk categories, 77 attack techniques, 30 environment categories, and 71 tool APIs.

Reviewer gioT

  • Clarified that our infrastructure goes beyond data integration through adaptive red-teaming and executable verification.
  • Reported verified attack success rates of 90.6% for direct attacks and 93.9% for combined direct and indirect attacks.
  • Explained the verifier-grounded labeling pipeline, which checks attack outcomes against MCP records and sandbox states, rejecting 7.30% of LLM-proposed successes.
  • Added controlled GRPO comparisons and component-wise ablations to establish the contribution of GuardPO.

Reviewer rtKT

  • Clarified GuardPO's decision-aligned optimization objective and its relationship to CISPO.
  • Added normalization ablations, showing accuracy drops of 3.26 and 9.55 percentage points when rationale and verdict normalization are removed, respectively.
  • Added reward-design ablations. The original reward configuration achieves the highest accuracy (90.88%) among the evaluated variants.
  • Explained the use of terminal outcome rewards and the limitations of dense rewards without verifiable process-level supervision.
  • Further detailed the contributions of adaptive red-teaming and executable verification.

Across all responses, we have:

  • Strengthened the validation of GuardPO through optimization and reward ablations.
  • Provided additional evidence of pre-execution safety detection.
  • Clarified the contributions of adaptive attack generation and execution-grounded supervision.
  • Reaffirmed our commitment to release model weights, training code, evaluation scripts, and relevant artifacts.

We believe these additional results and clarifications substantially address the reviewers' concerns and strengthen the rigor and practical relevance of HazardAuditor. We sincerely appreciate their valuable feedback.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.