{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.23189", "kind": "paper", "url": "https://arxiv.org/abs/2606.23189"}], "authors": ["Anmol Goel", "Iryna Gurevych"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.23189", "source_url": "https://arxiv.org/abs/2606.23189"}, "categories": ["General AI", "Language & Knowledge", "cs.AI"], "description": {"en": "AgentCIBench is an executable evaluation harness for testing whether computer-use agents improperly disclose information across application contexts."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.23189", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.23189", "title": "AgentCIBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2606.23189", "modality": null, "name": "AgentCIBench", "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_agentcibench_ed43589d", "origin_source": "arxiv", "recognition_confidence": "0.65", "record_sha256": "7848146ef892f188cdb713dbd1dee2f80ffeacf765af8d09aaf35583cbafa023", "relation": "introduces", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T17:43:54.508165Z", "source_url": "https://arxiv.org/abs/2606.23189"}, "publisher": null, "released": "2026-06-22", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.23189", "source_url": "https://arxiv.org/abs/2606.23189"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-23189", "source": "claire_radar", "source_benchmark_id": "2606.23189", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The benchmark addresses an immediate privacy risk for computer-use agents and reports failures across 15 frontier agents.", "score": 7}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T17:43:54.508165Z", "inputHash": "d35244fb5f4e32e5b7b98e2bee0b7791f77f1ed351f4da6d6efbd5ff0469ceb0", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "we introduce AgentCIBench, an evaluation harness", "canonicalNameSource": "abstract", "decisionReason": "The supplied abstract provides a formal name, executable task scenarios, deterministic scoring, and an explicit public-release statement.", "evaluationEvidence": {"reuse": "We release AgentCIBench", "scoring": "executable, deterministically scored scenarios", "task": "We target three common failure modes in CUAs: visual co-location"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T17:43:54.508165Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "AgentCIBench is an executable evaluation harness for testing whether computer-use agents improperly disclose information across application contexts.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["exact named benchmark artifact released in abstract", "evaluation protocol evidence"], "snippet": "Hence, we introduce AgentCIBench, an evaluation harness that turns this risk into executable, deterministically scored scenarios."}, "familyId": "bmf_f3973bc66534", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_agentcibench_ed43589d", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2606.23189", "pdf": "https://arxiv.org/pdf/2606.23189", "project": null, "report": "https://arxiv.org/abs/2606.23189"}, "metrics": [], "motivation": "Computer-use agents (CUAs) now act on a user's behalf across personal applications such as email, calendars, and to-do lists.", "name": "AgentCIBench", "oneLine": "AgentCIBench is an executable evaluation harness for testing whether computer-use agents improperly disclose information across application contexts.", "paperTitle": "Capable but Careless: Do Computer-Use Agents Follow Contextual Integrity?", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.23189", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "readiness": "Paper only", "recognitionConfidence": 0.65, "relation": "introduces", "releasedAt": "2026-06-22", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Anmol Goel and Iryna Gurevych"], "categories": ["cs.AI", "cs.CL"], "id": "2606.23189", "title": "Capable but Careless: Do Computer-Use Agents Follow Contextual Integrity?", "type": "arxiv", "url": "https://arxiv.org/abs/2606.23189"}, "sourceUpdatedAt": "2026-06-22T11:36:58Z", "topics": ["cs.AI"], "venueAttempts": [], "whyItMatters": "It provides deterministic testing of contextual-disclosure failures that can arise when agents operate across personal applications."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
