{"record": {"aliases": ["Domain-Conditioned Safety in Frontier Computer-Using Agents: A 793-Episode Browser Benchmark, a Coding-Domain Cross-Reference, and a Reproducibility Audit of Recent Red-Teaming"], "artifacts": [{"id": "arxiv:2606.05233", "kind": "paper", "url": "https://arxiv.org/abs/2606.05233"}], "authors": ["Nicholas Saban"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.05233", "source_url": "https://arxiv.org/abs/2606.05233"}, "categories": ["General AI", "Safety", "Safety & Trustworthiness"], "description": {"en": "A public 793-episode benchmark for evaluating prompt-injection attacks against computer-using agents across web tasks, attack templates, attack families, and system-prompt configurations."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.05233", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.05233", "title": "CUA-HandCrafted"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["University of California, Berkeley"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2606.05233", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2606.05233", "source_url": "https://openalex.org/W7163686832"}, "key": "claire-radar:2606.05233", "modality": null, "name": "CUA-HandCrafted", "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_domain-conditioned-safety-in-frontier-comp_f41845bd", "origin_source": "arxiv", "recognition_confidence": "0.6", "record_sha256": "a5730f3e2b513934ddfa9d820b9627210b7f84e7fb6f61134781b4f28708ff9b", "relation": "introduces", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T18:19:21.056978Z", "source_url": "https://arxiv.org/abs/2606.05233"}, "publisher": null, "released": "2026-06-03", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.05233", "source_url": "https://arxiv.org/abs/2606.05233"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-05233", "source": "claire_radar", "source_benchmark_id": "2606.05233", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["Domain-Conditioned Safety in Frontier Computer-Using Agents: A 793-Episode Browser Benchmark, a Coding-Domain Cross-Reference, and a Reproducibility Audit of Recent Red-Teaming"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Safety & Trustworthiness", "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The evaluation targets current frontier computer-using agents and reports a sharp browser-versus-coding safety contrast likely to attract near-term discussion.", "score": 7}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T18:19:21.056978Z", "inputHash": "ae855496e7ca0eece37c3955f8339521986293e7314fcb987fc557dafb66d51b", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "We release CUA-HandCrafted, a public benchmark", "canonicalNameSource": "abstract", "decisionReason": "The abstract formally names the benchmark, defines its task coverage, identifies attack success rate as the comparable score, and explicitly states that the benchmark is public.", "evaluationEvidence": {"reuse": "We release CUA-HandCrafted, a public benchmark", "scoring": "prompt-injection attack success rates (ASR)", "task": "793 episodes spanning 24 multi-step web tasks, 56 attack templates, 8 attack families, and 4 system-prompt configurations"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T18:19:21.056978Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "A public 793-episode benchmark for evaluating prompt-injection attacks against computer-using agents across web tasks, attack templates, attack families, and system-prompt configurations.", "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"], "snippet": "We release CUA-HandCrafted, a public benchmark of 793 episodes spanning 24 multi-step web tasks, 56 attack templates, 8 attack families, and 4 system-prompt configurations."}, "familyId": "bmf_342d34310686", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_domain-conditioned-safety-in-frontier-comp_f41845bd", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2606.05233", "pdf": "https://arxiv.org/pdf/2606.05233", "project": null, "report": "https://arxiv.org/abs/2606.05233"}, "metrics": [], "motivation": "Recent computer-using-agent (CUA) red-teaming papers report prompt-injection attack success rates (ASR) of 42-98%, but these headline numbers cluster on retired models and on the most-vulnerable model in each paper's panel.", "name": "CUA-HandCrafted", "oneLine": "A public 793-episode benchmark for evaluating prompt-injection attacks against computer-using agents across web tasks, attack templates, attack families, and system-prompt configurations.", "paperTitle": "Domain-Conditioned Safety in Frontier Computer-Using Agents: A 793-Episode Browser Benchmark, a Coding-Domain Cross-Reference, and a Reproducibility Audit of Recent Red-Teaming", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.05233", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "readiness": "Paper only", "recognitionConfidence": 0.6, "relation": "introduces", "releasedAt": "2026-06-03", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Nicholas Saban"], "categories": ["cs.CR", "cs.AI", "cs.CL"], "id": "2606.05233", "title": "Domain-Conditioned Safety in Frontier Computer-Using Agents: A 793-Episode Browser Benchmark, a Coding-Domain Cross-Reference, and a Reproducibility Audit of Recent Red-Teaming", "type": "arxiv", "url": "https://arxiv.org/abs/2606.05233"}, "sourceUpdatedAt": "2026-06-03T01:21:59Z", "topics": ["Safety"], "venueAttempts": [], "whyItMatters": "It provides a repeatable way to compare browser-agent attack success and examine whether safety behavior transfers across domains."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
