{"record": {"aliases": ["Understanding and Evaluating Claw-like Agent Security Through a Computer-Systems Lens"], "artifacts": [{"id": "arxiv:2606.30755", "kind": "paper", "url": "https://arxiv.org/abs/2606.30755"}, {"id": "gh:sunblaze-ucb/safeclawarena", "kind": "repo", "url": "https://github.com/sunblaze-ucb/SafeClawArena"}], "authors": ["Peizhi Niu", "Wenjie Qu", "Shangding Gu", "Tianneng Shi", "Yuankai Li", "Ahmad Tawaha", "Hend Alzahrani", "Vincent Siu", "Boyi Li", "Chenguang Wang", "Jiaheng Zhang", "Basel Alomair", "Ming Jin", "Muhao Chen", "Chi Wang", "Costas Spanos", "Dawn Song"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.30755", "source_url": "https://arxiv.org/abs/2606.30755"}, "categories": ["Agents", "Cybersecurity", "Language & Knowledge"], "description": {"en": "SafeClawArena evaluates security failures in Claw-like agents using 406 adversarial tasks spanning four attack surfaces in containerized platform replicas."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.30755", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.30755", "title": "SafeClawArena"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["University of Washington", "University of California, Berkeley", "Google (United States)"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2606.30755", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2606.30755", "source_url": "https://openalex.org/W7166871951"}, "key": "claire-radar:2606.30755", "modality": null, "name": "SafeClawArena", "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_understanding-and-evaluating-claw-like-age_e319f18b", "origin_source": "arxiv", "recognition_confidence": "0.7", "record_sha256": "da23f05f35f13d2dd9df98e370158d6b1dc6bf31808cbc248181394f04ff5ee3", "relation": "introduces", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T17:22:16.273226Z", "source_url": "https://arxiv.org/abs/2606.30755"}, "publisher": null, "released": "2026-06-29", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.30755", "source_url": "https://arxiv.org/abs/2606.30755"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-30755", "source": "claire_radar", "source_benchmark_id": "2606.30755", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["Understanding and Evaluating Claw-like Agent Security Through a Computer-Systems Lens"], "annotation": "Unknown", "applicationDomains": ["Cybersecurity"], "area": "Language & Knowledge", "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "Agent security, real-platform replication, and severe reported attack success rates may draw near-term technical attention.", "score": 74}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T17:22:16.273226Z", "inputHash": "099f5852d02ab5b60b3bec380cf21e290a3d055ea59992bf1d491f37c02003c6", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "we develop SafeClawArena", "canonicalNameSource": "abstract", "decisionReason": "Publish: the abstract names a repeatable benchmark, describes its automated evaluation mechanism, and supplies a public code-and-data path.", "evaluationEvidence": {"reuse": "Code and data: https://github.com/sunblaze-ucb/SafeClawArena.", "scoring": "evaluated via automated taint tracking across nine output channels", "task": "a benchmark of 406 adversarial tasks across four attack surfaces"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T17:22:16.273226Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "SafeClawArena evaluates security failures in Claw-like agents using 406 adversarial tasks spanning four attack surfaces in containerized platform replicas.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["benchmark term in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "From this perspective, we develop SafeClawArena, a benchmark of 406 adversarial tasks across four attack surfaces (Skill Supply-Chain Integrity, Persistent State Exploitation, Cross-Boundary Data Flow, and Indirect Prompt Injection), executed in containerized replicas of real agent platforms with canary-marked credentials and evaluated via automated taint tracking across nine output channels."}, "familyId": "bmf_ff0f2f845b95", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_understanding-and-evaluating-claw-like-age_e319f18b", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Cybersecurity"], "links": {"code": "https://github.com/sunblaze-ucb/SafeClawArena", "data": null, "paper": "https://arxiv.org/abs/2606.30755", "pdf": "https://arxiv.org/pdf/2606.30755", "project": null, "report": "https://arxiv.org/abs/2606.30755"}, "metrics": [], "motivation": "Claw-like AI agents (e.g., OpenClaw) are always-on processes with persistent access to credentials, files, tools, and external services.", "name": "SafeClawArena", "oneLine": "SafeClawArena evaluates security failures in Claw-like agents using 406 adversarial tasks spanning four attack surfaces in containerized platform replicas.", "paperTitle": "Understanding and Evaluating Claw-like Agent Security Through a Computer-Systems Lens", "primaryDomain": "Cybersecurity", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.30755", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "readiness": "Runnable", "recognitionConfidence": 0.7, "relation": "introduces", "releasedAt": "2026-06-29", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Peizhi Niu", "Wenjie Qu", "Shangding Gu", "Tianneng Shi", "Yuankai Li", "Ahmad Tawaha", "Hend Alzahrani", "Vincent Siu", "Boyi Li", "Chenguang Wang", "Jiaheng Zhang", "Basel Alomair", "Ming Jin", "Muhao Chen", "Chi Wang", "Costas Spanos", "Dawn Song"], "categories": ["cs.CR", "cs.AI"], "id": "2606.30755", "title": "Understanding and Evaluating Claw-like Agent Security Through a Computer-Systems Lens", "type": "arxiv", "url": "https://arxiv.org/abs/2606.30755"}, "sourceUpdatedAt": "2026-06-29T18:00:45Z", "topics": ["Agents"], "venueAttempts": [], "whyItMatters": "It measures cross-component security failures involving persistent state, supply chains, data flow, and indirect prompt injection."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
