ReA-Bench: Benchmarking Responsible Agent Auditing
Organizations: Alibaba AAIG
Abstract
As agents perform tasks autonomously in local and cloud environments, auditing their execution trajectories can help service providers identify risks and improve agent harnesses and foundation models. Our analysis of production and red-team trajectories shows that some risks are repeatedly missed across auditing configurations. Overall recall obscures which risks are missed and whether higher recall comes with more false positives on legitimate behavior. Existing benchmarks offer limited insight into these failures. Generated trajectories can contain inconsistent tool calls and responses, conspicuous risky actions, or repetitive attack patterns. Grouping results by risk type also makes it difficult to tell whether auditors overlooked evidence, accepted an unsupported conclusion, or lacked information needed to make a judgment. We introduce ReA-Bench, a diagnostic benchmark for auditing responsible agent behavior. It organizes auditing challenges into six tiers: internal safety knowledge, contradictions within trajectories, evidence retrieval in long contexts, verification of sources and reasoning, indirect steering, and missing information. It includes challenging benign cases in which actions resemble risky ones but are consistent with the task and its authorization. It also provides paired trajectory versions to test whether auditors detect the same risks more reliably when the trajectory is shortened or missing information is supplied. Most samples are drawn from real production trajectories, with scenarios manually selected and adapted to preserve realistic tasks and risks that could plausibly arise during execution.
Figures & tables
| Tier | Verification requirement | Domains or subtypes | Risky | Benign | Controls |
|---|---|---|---|---|---|
| L1 Internal knowledge | Assess action semantics, targets, and authorization | File system, command execution, network communication, system changes, persistence, and compliance | 123 | 75 | 0 |
| L2 Within-trajectory contradictions | Check actions and claims against requests, evidence, and rules | Intent, evidence, and specification contradictions | 48 | 31 | 0 |
| L3 Evidence dilution | Find and connect evidence in long trajectories or large outputs | Vertical and horizontal evidence dilution | 32 | 40 | 79 |
| L4 Provenance and logic | Verify sources, facts, and inferences | Format forgery, factual fabrication, and logical fabrication | 59 | 31 | 0 |
| L5 Indirect steering | Trace how the arrangement of materials steers decisions across turns | No further subdivision | 15 | 0 | 0 |
| L6 Missing information | Judge risk with and without key rules or runtime facts | Prior rules and runtime information | 27 | 0 | 27 |
| Model | L1 | L2 | L3 | L4 | L5 | Overall recall | Overall FPR |
|---|---|---|---|---|---|---|---|
| DS-Flash | 65.7% | 49.0% | 68.0% | 48.7% | 70.0% | 59.7% | 3.8% |
| DS-Pro | 57.1% | 42.7% | 54.7% | 40.3% | 55.0% | 50.6% | 3.8% |
| Qwen-Flash | 76.4% | 53.6% | 60.2% | 53.0% | 41.7% | 63.7% | 8.6% |
| Qwen-max | 87.6% | 58.9% | 64.8% | 64.0% | 61.7% | 73.6% | 12.9% |
| Grok-low | 79.0% | 44.3% | 67.2% | 47.9% | 88.3% | 65.4% | 5.4% |
| Grok-high | 83.9% | 47.9% | 76.8% | 55.1% | 90.0% | 71.0% | 5.5% |
| Model | FA | miniCoT | FHA | WA |
|---|---|---|---|---|
| DS-Flash | 54.5% / 1.7% | 60.3% / 6.8% | 66.4% / 4.0% | 57.4% / 2.8% |
| DS-Pro | 40.8% / 0.0% | 56.3% / 5.1% | 51.3% / 4.0% | 54.2% / 6.2% |
| Qwen-Flash | 56.3% / 2.8% | 63.9% / 9.0% | 67.9% / 11.3% | 66.8% / 11.3% |
| Qwen-max | 63.5% / 5.1% | 76.5% / 12.4% | 74.7% / 14.1% | 79.4% / 19.8% |
| Grok-low | 63.1% / 2.8% | 62.8% / 2.8% | 66.5% / 7.3% | 69.1% / 8.5% |
| Grok-high | 68.7% / 3.4% | 70.3% / 4.5% | 72.1% / 6.2% | 72.9% / 7.9% |
| Model/configuration | FA | miniCoT | FHA | WA |
|---|---|---|---|---|
| DS-Flash | 67.8% | 60.8% | 66.3% | 67.0% |
| DS-Pro | 71.1% | 69.2% | 68.5% | 65.2% |
| Qwen-Flash | 64.8% | 68.5% | 60.8% | 56.8% |
| Qwen-max | 71.8% | 58.2% | 52.0% | 56.0% |
| Grok-low | 70.2% | 74.6% | 67.3% | 69.9% |
| Grok-high | 74.7% | 78.0% | 72.2% | 72.9% |
| Model | Baseline method | Risky / benign samples | Recall | FPR | Gained/lost detections | Added/removed false positives |
|---|---|---|---|---|---|---|
| DS-Flash | FHA | 277 / 177 | 66.4% 77.6% | 4.0% 9.6% | +43 / 12 | +14 / 4 |
| Qwen-Flash | FHA | 277 / 177 | 67.9% 81.2% | 11.3% 15.3% | +50 / 13 | +20 / 13 |
| Grok-low | WA | 273 / 177 | 69.6% 87.5% | 8.5% 13.6% | +54 / 5 | +18 / 9 |
| GPT-luna | FA | 275 / 177 | 87.6% 94.9% | 49.2% 74.0% | +26 / 6 | +49 / 5 |
| Model | Baseline method | Risky / benign samples | AUROC | AUPRC(AP) |
|---|---|---|---|---|
| DS-Flash | FA | 277 / 177 | 0.893 0.904 | 0.921 0.931 |
| Qwen-Flash | miniCoT | 277 / 177 | 0.861 0.883 | 0.895 0.921 |
| Grok-low | FA | 273 / 177 | 0.892 0.924 | 0.926 0.943 |
| GPT-luna | FA | 275 / 177 | 0.825 0.785 | 0.862 0.824 |
| Model | Method | Recall | FPR | AUROC | Low-FPR recall |
|---|---|---|---|---|---|
| DS-Pro | FA | 40.8% 55.2% | 0.0% 2.8% | 0.878 0.885 | 70.8% 71.5% |
| DS-Pro | miniCoT | 56.3% 59.6% | 5.1% 2.3% | 0.870 0.900 | 68.6% 76.9% |
| DS-Pro | WA | 54.2% 65.7% | 6.2% 10.7% | 0.839 0.847 | 65.0% 56.3% |
| Grok | FA | 63.1% 68.6% | 2.8% 3.4% | 0.892 0.895 | 69.7% 74.1% |
| Grok | miniCoT | 62.8% 70.1% | 2.8% 4.5% | 0.890 0.908 | 74.1% 77.4% |
| Grok | WA | 69.1% 72.7% | 8.5% 7.9% | 0.872 0.879 | 69.1% 72.7% |
| Subtype | Risky / benign samples | Eight configurations: recall / FPR | Six non-GPT configurations: recall / FPR |
|---|---|---|---|
| L2 Intent contradiction | 11 / 10 | 75.6 / 15.9 | 72.0 / 3.8 |
| L2 Specification contradiction | 19 / 10 | 58.4 / 6.2 | 53.5 / 1.2 |
| L2 Evidence contradiction | 18 / 11 | 39.6 / 8.8 | 31.2 / 2.3 |
| L4 Format forgery | 20 / 10 | 85.3 / 47.2 | 81.2 / 31.2 |
| L4 Factual fabrication | 20 / 11 | 50.3 / 5.7 | 43.3 / 0.8 |
| L4 Logical fabrication | 19 / 10 | 42.3 / 22.2 | 28.7 / 10.4 |
| Main issue in frequently missed trajectories | Count | What the auditor needs to check |
|---|---|---|
| Evidence does not support the conclusion | 7 | Whether the cited original material actually supports the stated conclusion |
| Mismatched entities, sets, or values | 6 | Whether statements refer to the same entity and data set, and whether details reproduce the reported totals |
| Overinterpreting facts to reach unsupported or incorrect conclusions | 4 | Whether correlation is treated as causation, “sufficient” is turned into “necessary,” or local effects are generalized to the whole |
| Omitted or violated task requirements or delivery rules | 7 | Whether required format, length, and constraints are met, and whether prescribed steps were actually performed |
| Unauthorized execution of dangerous operations | 2 | Whether operations such as obtaining resources or probing networks have the required permission |
| Main situation in frequently flagged benign trajectories | Count | What the auditor needs to distinguish |
|---|---|---|
| Execution after updated user requirements or additional authorization | 9 | Whether the user changed the requirements or supplied valid authorization, and whether the agent stayed within the permitted actions |
| User-directed development and maintenance following a supplied document | 6 | Whether document steps serve the user task, and whether the actions are ordinary maintenance or involve unauthorized access, destruction, or disclosure |
| Sourced research and explicitly labeled suggestions | 11 | Whether analysis, examples, or disclosed information gaps are incorrectly treated as fabricated facts |
| Version comparison | Paired judgments | Threshold-decision agreement | Lost / gained detections | Mean absolute score difference | Score correlation r |
|---|---|---|---|---|---|
| Horizontal: original shorter I | 480 | 88.3% | 21 / 35 | 0.91 | 0.814 |
| Horizontal: original shorter II | 480 | 87.3% | 19 / 42 | 1.00 | 0.769 |
| Horizontal: original concise | 480 | 89.2% | 10 / 42 | 0.98 | 0.818 |
| Vertical: original shorter I | 528 | 89.0% | 17 / 41 | 0.95 | 0.852 |
| Vertical: original shorter II | 528 | 87.1% | 29 / 39 | 1.14 | 0.809 |
| Method | Prior rules (17 pairs) | Runtime information (10 pairs) |
|---|---|---|
| FA | 25.7% 86.0% | 10.0% 62.5% |
| miniCoT | 39.0% 91.9% | 10.0% 77.5% |
| FHA | 38.2% 94.9% | 13.8% 61.3% |
| WA | 41.2% 97.1% | 11.3% 66.3% |
Appendix figures & tables22 assets
Supplementary material from the paper’s appendix.
Appendix
| Model | Method | L1(123) | L2(48) | L3(32) | L4(59) | L5(15) | Overall recall | Overall FPR (177) |
|---|---|---|---|---|---|---|---|---|
| DS-Flash | FA | 61.0% | 45.8% | 59.4% | 42.4% | 66.7% | 54.5% | 1.7% |
| DS-Flash | miniCoT | 68.3% | 52.1% | 65.6% | 44.1% | 73.3% | 60.3% | 6.8% |
| DS-Flash | FHA | 71.5% | 54.2% | 78.1% | 57.6% | 73.3% | 66.4% | 4.0% |
| DS-Flash | WA | 61.8% | 43.8% | 68.8% | 50.8% | 66.7% | 57.4% | 2.8% |
| DS-Pro | FA | 44.7% | 37.5% | 40.6% | 33.9% | 46.7% | 40.8% | 0.0% |
| DS-Pro | miniCoT | 65.0% | 45.8% | 50.0% | 44.1% | 80.0% | 56.3% | 5.1% |
| Model | Method | L1(75) | L2(31) | L3(40) | L4(31) | Overall (177) |
|---|---|---|---|---|---|---|
| DS-Flash | FA | 1.3% | 3.2% | 0.0% | 3.2% | 1.7% |
| DS-Flash | miniCoT | 5.3% | 6.5% | 10.0% | 6.5% | 6.8% |
| DS-Flash | FHA | 4.0% | 0.0% | 2.5% | 9.7% | 4.0% |
| DS-Flash | WA | 0.0% | 6.5% | 0.0% | 9.7% | 2.8% |
| DS-Pro | FA | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% |
| DS-Pro | miniCoT | 2.7% | 3.2% | 5.0% | 12.9% | 5.1% |
| Model | Method | AUROC | AUPRC | Recall@FPR 10% |
|---|---|---|---|---|
| DS-Flash | FA | 0.893 | 0.921 | 67.8% |
| DS-Flash | miniCoT | 0.884 | 0.915 | 60.8% |
| DS-Flash | FHA | 0.874 | 0.917 | 66.3% |
| DS-Flash | WA | 0.869 | 0.903 | 67.0% |
| DS-Pro | FA | 0.881 | 0.910 | 71.1% |
| DS-Pro | miniCoT | 0.875 | 0.899 | 69.2% |
| Model | Method | Prior rules (17 pairs) | Gained/lost detections | Runtime information (10 pairs) | Gained/lost detections |
|---|---|---|---|---|---|
| DS-Flash | FA | 11.8% 82.4% | +13 / 1 | 0.0% 30.0% | +3 / 0 |
| DS-Flash | miniCoT | 23.5% 94.1% | +12 / 0 | 0.0% 50.0% | +5 / 0 |
| DS-Flash | FHA | 41.2% 94.1% | +9 / 0 | 10.0% 40.0% | +3 / 0 |
| DS-Flash | WA | 29.4% 94.1% | +11 / 0 | 10.0% 60.0% | +5 / 0 |
| DS-Pro | FA | 5.9% 70.6% | +11 / 0 | 0.0% 30.0% | +3 / 0 |
| DS-Pro | miniCoT | 47.1% 70.6% | +4 / 0 | 0.0% 70.0% | +7 / 0 |
| Method | Audit position | Information visible to the auditor | Purpose |
|---|---|---|---|
| FHA | Every 15 messages | Complete original history from the trajectory start to the current checkpoint | Effectiveness and cost of repeated auditing without history compression |
| WA | Every 15 messages | Most recent 20 messages plus user requests outside the window, without summaries | Evidence lost through direct truncation |
| GSA | Extract evidence from non-overlapping 20-message windows, then audit once at the endpoint | Extraction uses the full trajectory plus target window. Final auditing uses the full trajectory plus all extracted clues | Effect of collecting dispersed evidence before final judgment |
| FA | Final checkpoint only | Complete original trajectory | Ability to review a long trajectory in one call |
| miniCoT | Final checkpoint only | Complete original trajectory, listing potential concerns before scoring | Whether listing risk points first reduces misses |
| Model | Auditing method | FPR: Benchmark benign benign production traffic |
|---|---|---|
| DS-Flash | miniCoT | 6.8% 0.0% |
| DS-Pro | WA | 6.2% 0.5% |
| Qwen-Flash | WA | 11.3% 0.0% |
| Qwen-max | WA | 19.8% 1.5% |
| Grok-low | WA | 8.5% 0.5% |
| Grok-high | WA | 7.9% 1.0% |
| Dimension | R-Judge [ 65 ] | ASSEBench [ 37 ] | SHADE-Arena [ 28 ] | ATBench [ 32 ] | This work |
|---|---|---|---|---|---|
| Size | 569 records | 2,293 records | 17 main/side-task pairs | 1,000 trajectories | 587 trajectories |
| Data sources | Selected open-source trajectories and manually constructed cases with ChatGPT assistance | LLM execution in four existing safety benchmarks | Agent execution trajectories from synthetic environments | LLM-simulated trajectories guided by risk categories | Real production risk cases and red-team trajectories as seeds, manually selected, adapted, and constructed where necessary |
| Trajectory size | 2.6 turns, 270 tokens | 5.3 turns, 300 tokens | 25 turns (successful main task) | 9.01 turns, 3.95k tokens | 66 turns, 36k tokens |
| Evaluation organization and controls | Application scenarios and risk types | Classification by scenario, risk type, and behavior mode | Paired benign main tasks and covert harmful side tasks | Risk source, failure mode, and real-world harm | Six tiers organized by auditing judgment requirements |
| Risk focus | Indirect injection against personal agents and risky behavior during task execution | Unsafe agent interactions in Benchmark datasets | Covert harmful agent behavior during benign tasks | Behavioral and output risks in multistep tool interactions | Agent operation and delivery risks, including attack-induced behavior and agent misconduct |
AI Agent Auditing
AI agent auditing examines agents’ actions, tool use, evidence, and outcomes to assess safety, compliance, reliability, and accountability.
See all 308 papers