{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.09476", "kind": "paper", "url": "https://arxiv.org/abs/2608.09476"}, {"id": "gh:zjuicsr/actbench", "kind": "repo", "url": "https://github.com/zjuicsr/ActBench"}], "authors": ["Hongwei Yao", "Yiming Liu", "Meihui Chen", "Jieling Chen", "Zikun Chen", "Yiling He", "Wangze Ni", "Cong Wang", "Kui Ren"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.09476", "source_url": "https://arxiv.org/abs/2608.09476"}, "categories": ["General AI", "Safety", "Safety & Trustworthiness"], "description": {"en": "Self-evolving benchmark of behavioral safety in cowork agents, evaluating risk from execution trajectories across 600 cases, 213 scenarios, 15 risk behaviors, and 48 web-service APIs."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.09476", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.09476", "title": "ActBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["City University of Hong Kong", "Zhejiang University", "University College London"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2608.09476", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2608.09476", "source_url": "https://openalex.org/W7202180668"}, "key": "claire-radar:2608.09476", "modality": null, "name": "ActBench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_actbench_33a6cb88", "origin_source": "arxiv", "recognition_confidence": "0.75", "record_sha256": "5821b86f2d7a155095121c8718245962d2877e89a44da367e6f1fe969b673c1d", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-09-02T08:54:07.291235Z", "source_url": "https://arxiv.org/abs/2608.09476"}, "publisher": null, "released": "2026-08-10", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.09476", "source_url": "https://arxiv.org/abs/2608.09476"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-09476", "source": "claire_radar", "source_benchmark_id": "2608.09476", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Safety & Trustworthiness", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": "https://github.com/zjuicsr/ActBench", "githubScope": "benchmark_repo", "githubStars": 6, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 1, "hfPaperUrl": "https://huggingface.co/papers/2608.09476", "observedAt": "2026-10-04T15:31:48.030916Z", "signalStatus": {"githubStars": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-10-01", "state": "stale"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}}, "status": "fresh"}, "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The benchmark addresses a timely topic in agent safety with a large trajectory dataset and public tooling, likely attracting attention from AI safety and agent evaluation communities.", "score": 70}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-02T08:54:07.291235Z", "inputHash": "800d51a23ee4c687e7398cc22e9ee0450c0a8539d93e53f053e00065d0e1d50c", "model": "deepseek-v4-pro"}, "curation": {"canonicalNameEvidence": "The project and benchmark are named **ActBench** and live in `ZJUICSR/ActBench`.", "canonicalNameSource": "official_readme", "decisionReason": "Formally named benchmark with released cases, trajectory runner, scorers, and public Hugging Face dataset; includes stable scoring metrics.", "model": "deepseek-v4-pro", "reviewedAt": "2026-09-02T08:54:07.291235Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Self-evolving benchmark of behavioral safety in cowork agents, evaluating risk from execution trajectories across 600 cases, 213 scenarios, 15 risk behaviors, and 48 web-service APIs.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "score_submission", "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "evaluation protocol evidence", "public artifact URL"], "snippet": "We define behavioral safety and introduce ActBench, a self-evolving benchmark that evaluates such behavior risk from execution trajectories rather than final responses."}, "familyId": "bmf_9c06921b8b0c", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_actbench_33a6cb88", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/zjuicsr/ActBench", "data": null, "paper": "https://arxiv.org/abs/2608.09476", "pdf": "https://arxiv.org/pdf/2608.09476", "project": null, "report": "https://arxiv.org/abs/2608.09476"}, "metrics": [], "motivation": "Cowork agents may complete benign tasks while disclosing protected data, manipulating unauthorized state, invocate unauthorized API.", "name": "ActBench", "oneLine": "Self-evolving benchmark of behavioral safety in cowork agents, evaluating risk from execution trajectories across 600 cases, 213 scenarios, 15 risk behaviors, and 48 web-service APIs.", "paperTitle": "ActBench: Self-Evolving Benchmark of Behavioral Safety in Cowork Agents", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2608.09476", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "publishers": [{"name": "Zhejiang University ICSR", "organizationType": "academic-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/zjuicsr/ActBench"}], "ranking": {"90d": {"confidence": "Medium", "coverage": 0.7, "growth": {"components": {"githubStars": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 90}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": 0.2597440249823917, "value": 6}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": 0.12247537038455211, "value": 1}}, "confidence": "Medium", "coverage": 0.7, "rank": 235, "score": 31}, "method": "attention-ranking-v13", "rank": 235, "score": 31}}, "readiness": "Runnable", "recognitionConfidence": 0.75, "relation": "introduces", "releasedAt": "2026-08-10", "source": {"authors": ["Hongwei Yao", "Yiming Liu", "Meihui Chen", "Jieling Chen", "Zikun Chen", "Yiling He", "Wangze Ni", "Cong Wang", "Kui Ren"], "categories": ["cs.CR", "cs.AI"], "id": "2608.09476", "title": "ActBench: Self-Evolving Benchmark of Behavioral Safety in Cowork Agents", "type": "arxiv", "url": "https://arxiv.org/abs/2608.09476"}, "sourceUpdatedAt": "2026-08-10T11:45:03Z", "topics": ["Safety"], "venueAttempts": [], "whyItMatters": "Provides a standardized assessment of behavioral safety beyond final responses, enabling comparison of agent robustness against task-reachable attacks."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
