{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2607.01793", "kind": "paper", "url": "https://arxiv.org/abs/2607.01793"}, {"id": "gh:yunhao-feng/vera", "kind": "repo", "url": "https://github.com/Yunhao-Feng/Vera"}], "authors": ["Yunhao Feng", "Ruixiao Lin", "Ming Wen", "Qinqin He", "Yanming Guo", "Yifan Ding", "Yutao Wu", "Jialuo Chen", "Zhuoer Xu", "Xiaohu Du", "Jianan Ma", "Zixing Chen", "Xingjun Ma", "Yunhao Chen", "Xinhao Deng"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2607.01793", "source_url": "https://arxiv.org/abs/2607.01793"}, "categories": ["Agents", "Safety", "Safety & Trustworthiness", "Software & AI Compute"], "description": {"en": "Vera-Bench is an executable safety benchmark for tool-using agents, with deterministic, artifact-grounded verification across diverse risks and execution settings."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2607.01793", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2607.01793", "title": "Vera-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Zhejiang University", "Fudan University", "Alibaba Group (China)", "Deakin University"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2607.01793", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2607.01793", "source_url": "https://openalex.org/W7167202790"}, "key": "claire-radar:2607.01793", "modality": null, "name": "Vera-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_vera-bench_8e56a1ff", "origin_source": "arxiv", "recognition_confidence": "0.6", "record_sha256": "bbf1779b27db42358ebffe9531cf506480cc34c3d018ab95a23f3fc506e8172d", "relation": "introduces", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T19:01:37.943600Z", "source_url": "https://arxiv.org/abs/2607.01793"}, "publisher": null, "released": "2026-07-02", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2607.01793", "source_url": "https://arxiv.org/abs/2607.01793"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2607-01793", "source": "claire_radar", "source_benchmark_id": "2607.01793", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["Software & AI Compute"], "area": "Safety & Trustworthiness", "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The large executable case suite, public code, and evaluation of multiple production agent frameworks provide a strong basis for near-term technical interest.", "score": 82}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T18:59:59.467212Z", "inputHash": "0270091a78896e6801f0808b319eafaf7d8a91b1cf6026dcfb13d63234c3284f", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "we also release Vera-Bench", "canonicalNameSource": "abstract", "decisionReason": "The supplied abstract establishes a formal name, executable task suite, deterministic scoring contract, and public code path.", "evaluationEvidence": {"reuse": "The code is publicly available at https://github.com/Yunhao-Feng/Vera.", "scoring": "a deterministic verification predicate grounded in observable artifacts", "task": "comprising 1600 executable safety cases spanning 124 risk categories across three execution settings"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T19:01:37.943600Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Vera-Bench is an executable safety benchmark for tool-using agents, with deterministic, artifact-grounded verification across diverse risks and execution settings.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["exact named benchmark artifact released in abstract", "public artifact URL"], "snippet": "We evaluate Vera on four production agent frameworks (OpenClaw, Hermes, Codex, Claude Code), revealing substantial safety weaknesses, with average attack success rates reaching 93.9\\% under multi-channel attacks; we also release Vera-Bench, comprising 1600 executable safety cases spanning 124 risk categories across three execution settings."}, "familyId": "bmf_e922d191191c", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_vera-bench_8e56a1ff", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Software & Cloud"], "links": {"code": "https://github.com/Yunhao-Feng/Vera", "data": null, "paper": "https://arxiv.org/abs/2607.01793", "pdf": "https://arxiv.org/pdf/2607.01793", "project": null, "report": "https://arxiv.org/abs/2607.01793"}, "metrics": [], "motivation": "LLM agents increasingly perform autonomous actions through external tools, leading to complex and evolving safety risks.", "name": "Vera-Bench", "oneLine": "Vera-Bench is an executable safety benchmark for tool-using agents, with deterministic, artifact-grounded verification across diverse risks and execution settings.", "paperTitle": "Safety Testing LLM Agents at Scale: From Risk Discovery to Evidence-Grounded Verification", "primaryDomain": "Software & AI Compute", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2607.01793", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "readiness": "Runnable", "recognitionConfidence": 0.6, "relation": "introduces", "releasedAt": "2026-07-02", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Yunhao Feng", "Ruixiao Lin", "Ming Wen", "Qinqin He", "Yanming Guo", "Yifan Ding", "Yutao Wu", "Jialuo Chen", "Zhuoer Xu", "Xiaohu Du", "Jianan Ma", "Zixing Chen", "Xingjun Ma", "Yunhao Chen", "Xinhao Deng"], "categories": ["cs.AI"], "id": "2607.01793", "title": "Safety Testing LLM Agents at Scale: From Risk Discovery to Evidence-Grounded Verification", "type": "arxiv", "url": "https://arxiv.org/abs/2607.01793"}, "sourceUpdatedAt": "2026-07-04T02:16:40Z", "topics": ["Agents", "Safety"], "venueAttempts": [], "whyItMatters": "Its executable cases and environment-grounded predicates support scalable, comparable safety testing without relying on model self-report."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
