{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2605.22321", "kind": "paper", "url": "https://arxiv.org/abs/2605.22321"}, {"id": "gh:antgroup/agent3sigma-stage", "kind": "repo", "url": "https://github.com/antgroup/Agent3Sigma-Stage"}], "categories": ["Agents", "Agents & Tool Use", "General AI"], "description": {"en": "A3S-Bench is an executable security test suite for evaluating autonomous agents across multi-turn scenarios and agent-level risks using an action-grounded oracle."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2605.22321", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2605.22321", "title": "A3S-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2605.22321", "modality": null, "name": "A3S-Bench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"admission_policy_version": "2026-09-09.1", "confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-09-25T04:40:22Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_a3s-bench_370234a8", "origin_source": "arxiv", "recognition_confidence": "0.55", "record_sha256": "4797cd64e87e671ae6ba4722fdb016ea7a2d3ed9de0e18a8666aac1a6b346e1a", "relation": "evaluates_only", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T19:39:58.745325Z", "source_url": "https://arxiv.org/abs/2605.22321"}, "publisher": null, "released": "2026-05-21", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2605.22321", "source_url": "https://arxiv.org/abs/2605.22321"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2605-22321", "source": "claire_radar", "source_benchmark_id": "2605.22321", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Agents & Tool Use", "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The benchmark addresses autonomous-agent security with executable cases and a public implementation, but near-term attention cannot be established from the supplied source alone.", "score": 3}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T19:35:17.646109Z", "inputHash": "3017dec5a57327057436c7ff7aa07c5ef09e6a8038a2d921b8c57183fe6d27f4", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "We instantiate ASEval to construct A3S-Bench, a test suite of 2,254 executable multi-turn test cases", "canonicalNameSource": "abstract", "decisionReason": "Publish because the source formally names A3S-Bench, defines executable test cases and an action-grounded risk-trigger evaluation, and provides a currently available public code-and-data link.", "evaluationEvidence": {"reuse": "Code and data available at https://github.com/antgroup/Agent3Sigma-Stage", "scoring": "average risk-trigger rate", "task": "executes each case against the target agent"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T19:39:58.745325Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "A3S-Bench is an executable security test suite for evaluating autonomous agents across multi-turn scenarios and agent-level risks using an action-grounded oracle.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["exact named benchmark artifact released in abstract", "evaluation protocol evidence", "public artifact URL", "named entity is a model, method, framework, or dataset used in evaluation"], "snippet": "We instantiate ASEval to construct A3S-Bench, a test suite of 2,254 executable multi-turn test cases spanning diverse usage scenarios, delivery patterns, and agent-level security risks."}, "familyId": "bmf_fad6dcbfd267", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_a3s-bench_370234a8", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/antgroup/Agent3Sigma-Stage", "data": null, "paper": "https://arxiv.org/abs/2605.22321", "pdf": "https://arxiv.org/pdf/2605.22321", "project": null, "report": "https://arxiv.org/abs/2605.22321"}, "metrics": [], "motivation": "As autonomous agents (e.g., OpenClaw) increasingly operate with deep system-level privileges to execute complex tasks, they introduce severe, unmitigated security risks.", "name": "A3S-Bench", "oneLine": "A3S-Bench is an executable security test suite for evaluating autonomous agents across multi-turn scenarios and agent-level risks using an action-grounded oracle.", "paperTitle": "ASEval: Automated Trajectory-Level Security Testing for Autonomous Agents", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2605.22321", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "readiness": "Runnable", "recognitionConfidence": 0.55, "relation": "evaluates_only", "releasedAt": "2026-05-21", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Jianan Ma", "Xiaohu Du", "Ruixiao Lin", "Yaoxiang Bian", "Jialuo Chen", "Yunhao Feng", "Xiaofang Yang", "Shiwen Cui", "Changhua Meng", "Xinhao Deng", "Jingyi Wang", "Zhen Wang"], "categories": ["cs.CR", "cs.AI", "cs.SE"], "id": "2605.22321", "title": "ASEval: Automated Trajectory-Level Security Testing for Autonomous Agents", "type": "arxiv", "url": "https://arxiv.org/abs/2605.22321"}, "sourceUpdatedAt": "2026-07-24T06:09:29Z", "topics": ["Agents"], "venueAttempts": [], "whyItMatters": "It evaluates security failures across complete agent trajectories, including persistent state and tool-mediated actions that response-level testing can overlook."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
