{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.13994", "kind": "paper", "url": "https://arxiv.org/abs/2606.13994"}, {"id": "hf:decompositionbench/decompbench", "kind": "dataset", "url": "https://huggingface.co/datasets/decompositionbench/DeCompBench"}], "authors": ["Vikhyath Kothamasu", "Virginia Smith", "Chhavi Yadav"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.13994", "source_url": "https://arxiv.org/abs/2606.13994"}, "categories": ["Agents", "General AI", "Safety", "Safety & Trustworthiness"], "description": {"en": "DeCompBench evaluates agent safety against decomposition attacks, where harmful tasks are broken into benign subtasks; it measures refusal rates and objective fulfillment on decomposed variants."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.13994", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.13994", "title": "DeCompBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Carnegie Mellon University", "University of California, Berkeley"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2606.13994", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2606.13994", "source_url": "https://openalex.org/W7164844745"}, "key": "claire-radar:2606.13994", "modality": null, "name": "DeCompBench", "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_decompbench_f0f37f4f", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "cb42159a8c3e7199bc83de1642a007bc11f80be4b475856a134811062f95ae0f", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2606.13994"}, "publisher": null, "released": "2026-06-12", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.13994", "source_url": "https://arxiv.org/abs/2606.13994"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-13994", "source": "claire_radar", "source_benchmark_id": "2606.13994", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Safety & Trustworthiness", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": "decompositionbench/DeCompBench", "hfDatasetDownloads": 42, "hfDatasetLikes": 4, "hfPaperUpvotes": 1, "hfPaperUrl": "https://huggingface.co/papers/2606.13994", "observedAt": "2026-10-04T03:02:42.754940Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-10-01", "state": "stale"}, "hfDatasetLikes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-10-01", "state": "stale"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-10-01", "state": "stale"}}, "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": "13f49fa85a55bff0af3e7ab967aab671bf9160d72b2b6ee6fbd1ecf49e0a55aa", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "DeCompBench evaluates agent safety against decomposition attacks, where harmful tasks are broken into benign subtasks; it measures refusal rates and objective fulfillment on decomposed variants.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact named benchmark artifact released in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "To this end, we introduce DeCompBench, a benchmark designed specifically to evaluate agentic safety under decomposition attacks."}, "familyId": "bmf_820b856726b1", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_decompbench_f0f37f4f", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": "https://huggingface.co/datasets/decompositionbench/DeCompBench", "hfPaper": "https://huggingface.co/papers/2606.13994", "paper": "https://arxiv.org/abs/2606.13994", "pdf": "https://arxiv.org/pdf/2606.13994", "project": null, "report": "https://arxiv.org/abs/2606.13994"}, "metrics": [], "motivation": "LLM-based Agents are becoming increasingly capable and widely deployed, creating growing incentives for adversarial misuse in the real-world.", "name": "DeCompBench", "oneLine": "DeCompBench evaluates agent safety against decomposition attacks, where harmful tasks are broken into benign subtasks; it measures refusal rates and objective fulfillment on decomposed variants.", "paperTitle": "Hidden in Plain Sight: Benchmarking Agent Safety Against Decomposition Attacks with DECOMPBENCH", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.13994", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "ranking": {}, "readiness": "Inspectable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-06-12", "source": {"authors": ["Vikhyath Kothamasu", "Virginia Smith", "Chhavi Yadav"], "categories": ["cs.CR", "cs.AI", "cs.LG"], "id": "2606.13994", "title": "Hidden in Plain Sight: Benchmarking Agent Safety Against Decomposition Attacks with DECOMPBENCH", "type": "arxiv", "url": "https://arxiv.org/abs/2606.13994"}, "sourceUpdatedAt": "2026-06-12T00:30:29Z", "topics": ["Agents", "Safety"], "venueAttempts": [], "whyItMatters": "Addresses a security gap not covered by existing agent safety benchmarks, critical for preventing adversarial misuse in deployed agents."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
