{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.18356", "kind": "paper", "url": "https://arxiv.org/abs/2606.18356"}, {"id": "hf:sairights/safeclawbench", "kind": "dataset", "url": "https://huggingface.co/datasets/sairights/safeclawbench"}], "authors": ["Yuchuan Tian", "Mengyu Zheng", "Haocheng Mei", "Ye Yuan", "Chao Xu", "Xinghao Chen", "Hanting Chen", "Yu Wang"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.18356", "source_url": "https://arxiv.org/abs/2606.18356"}, "categories": ["Agents", "General AI", "Language & Knowledge"], "description": {"en": "SafeClawBench is a staged benchmark for tool-using LLM agent security with 600 adversarial tasks across six attack families, reporting three endpoints: semantic attack acceptance, audit-visible harm evidence, and sandbox-observed tool/state harm."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.18356", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.18356", "title": "SafeClawBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Peking University", "Beijing Jiaotong University", "Huawei Technologies (China)", "Tsinghua University"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2606.18356", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2606.18356", "source_url": "https://openalex.org/W7165183485"}, "key": "claire-radar:2606.18356", "modality": null, "name": "SafeClawBench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "High", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "unknown", "origin_record_id": "bm_safeclawbench_e7a55be8", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "38afb1b46b795a7731f62c45ab17843a70e48c0d0c6610846920d2d7ef3b4578", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2606.18356"}, "publisher": null, "released": "2026-06-16", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.18356", "source_url": "https://arxiv.org/abs/2606.18356"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-18356", "source": "claire_radar", "source_benchmark_id": "2606.18356", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": "sairights/safeclawbench", "hfDatasetDownloads": 267, "hfDatasetLikes": 0, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2606.18356", "observedAt": "2026-10-04T03:02:42.754940Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-10-02", "state": "stale"}, "hfDatasetLikes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-10-02", "state": "stale"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "state": "unavailable"}}, "status": "stale"}, "capabilities": [], "confidence": "High", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "1e8955f32543d95d9f59dfad38699b2ed8ce8a462bccecdf4748be32ffca13c4", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "SafeClawBench is a staged benchmark for tool-using LLM agent security with 600 adversarial tasks across six attack families, reporting three endpoints: semantic attack acceptance, audit-visible harm evidence, and sandbox-observed tool/state harm.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "We introduce SafeClawBench, a staged benchmark for tool-using agent security with 600 controlled adversarial tasks across six attack families: direct and indirect prompt injection, tool-return injection, memory poisoning, memory extraction, and ambiguity-driven unsafe inference."}, "familyId": "bmf_91d28b9baddb", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_safeclawbench_e7a55be8", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": "https://huggingface.co/datasets/sairights/safeclawbench", "hfPaper": "https://huggingface.co/papers/2606.18356", "paper": "https://arxiv.org/abs/2606.18356", "pdf": "https://arxiv.org/pdf/2606.18356", "project": null, "report": "https://arxiv.org/abs/2606.18356"}, "metrics": [], "motivation": "Tool-using language-model agents introduce security failures that go beyond unsafe text: they can disclose protected objects, write persistent memory, send messages, modify databases, or trigger harmful code and tool effects.", "name": "SafeClawBench", "oneLine": "SafeClawBench is a staged benchmark for tool-using LLM agent security with 600 adversarial tasks across six attack families, reporting three endpoints: semantic attack acceptance, audit-visible harm evidence, and sandbox-observed tool/state harm.", "paperTitle": "SafeClawBench: Separating Semantic, Audit-Evidence, and Sandbox Harm in Tool-Using LLM Agents", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.18356", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "SAIRights", "organizationType": "company-research-lab", "role": "benchmark-publisher", "sourceUrl": "https://huggingface.co/datasets/sairights/safeclawbench"}], "ranking": {}, "readiness": "Inspectable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-06-16", "source": {"authors": ["Yuchuan Tian", "Mengyu Zheng", "Haocheng Mei", "Ye Yuan", "Chao Xu", "Xinghao Chen", "Hanting Chen", "Yu Wang"], "categories": ["cs.CR", "cs.AI"], "id": "2606.18356", "title": "SafeClawBench: Separating Semantic, Audit-Evidence, and Sandbox Harm in Tool-Using LLM Agents", "type": "arxiv", "url": "https://arxiv.org/abs/2606.18356"}, "sourceUpdatedAt": "2026-06-16T18:04:45Z", "topics": ["Agents"], "venueAttempts": [], "whyItMatters": "SafeClawBench addresses the evaluation gap where existing benchmarks collapse distinct security failure stages into a single metric, making it hard to distinguish semantic compliance from actual harm. It provides separate, comparable scores across models and prompt policies, aiding in selecting agents and defenses based on the specific type of security risk."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
