{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2605.29224", "kind": "paper", "url": "https://arxiv.org/abs/2605.29224"}], "authors": ["Aditya Nawal", "Manit Baser", "Mohan Gurusamy"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2605.29224", "source_url": "https://arxiv.org/abs/2605.29224"}, "categories": ["Cybersecurity", "Information retrieval", "Safety", "Safety & Trustworthiness"], "description": {"en": "AgentREVEAL is a diagnostic framework that evaluates safety alignment degradation in LLM agents when web retrieval is integrated. It assesses the impact of retrieval integration and content properties on harmful compliance, using a set of harmful behaviors and retrieval sources."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2605.29224", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2605.29224", "title": "HarmURLBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2605.29224", "modality": null, "name": "HarmURLBench", "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_harmurlbench_cbd88280", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "88893d0256cb71ba7e9ff7278f73061a802ab73efd8d24556a03d80cc3e4f4a8", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2605.29224"}, "publisher": null, "released": "2026-05-28", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2605.29224", "source_url": "https://arxiv.org/abs/2605.29224"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2605-29224", "source": "claire_radar", "source_benchmark_id": "2605.29224", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["Cybersecurity"], "area": "Safety & Trustworthiness", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2605.29224", "observedAt": "2026-10-04T03:02:42.754940Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "state": "unavailable"}}, "status": "unavailable"}, "capabilities": ["Information retrieval"], "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": "52ef7800702e2fbae6e4cb6a8150cc4b1cba78efbbe66c98874fe2e988564532", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "AgentREVEAL is a diagnostic framework that evaluates safety alignment degradation in LLM agents when web retrieval is integrated. It assesses the impact of retrieval integration and content properties on harmful compliance, using a set of harmful behaviors and retrieval sources.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact named benchmark artifact released in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "We introduce HarmURLBench, a benchmark containing 1,405 real-world URLs paired with 320 harmful behaviors to support future evaluations."}, "familyId": "bmf_2962ddda58cf", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_harmurlbench_cbd88280", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Cybersecurity"], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2605.29224", "paper": "https://arxiv.org/abs/2605.29224", "pdf": "https://arxiv.org/pdf/2605.29224", "project": null, "report": "https://arxiv.org/abs/2605.29224"}, "metrics": [], "motivation": "AI agents augment large language models with external tools such as web retrieval, enabling grounded and up-to-date responses.", "name": "HarmURLBench", "oneLine": "AgentREVEAL is a diagnostic framework that evaluates safety alignment degradation in LLM agents when web retrieval is integrated. It assesses the impact of retrieval integration and content properties on harmful compliance, using a set of harmful behaviors and retrieval sources.", "paperTitle": "Relevance as a Vulnerability: How Web Retrieval Degrades Safety Alignment in LLM Agents", "primaryDomain": "Cybersecurity", "publication": {"evidence": "Accepted to EMNLP 2026 Findings", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2605.29224", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "EMNLP 2026 Findings", "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "arXiv", "organizationType": "community", "role": "benchmark-publisher", "sourceUrl": "https://arxiv.org/abs/2605.29224"}], "ranking": {}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-05-28", "source": {"authors": ["Aditya Nawal", "Manit Baser", "Mohan Gurusamy"], "categories": ["cs.CL", "cs.AI", "cs.CR"], "id": "2605.29224", "title": "Relevance as a Vulnerability: How Web Retrieval Degrades Safety Alignment in LLM Agents", "type": "arxiv", "url": "https://arxiv.org/abs/2605.29224"}, "sourceUpdatedAt": "2026-05-28T01:23:48Z", "topics": ["Safety"], "venueAttempts": [{"decisionRaw": "Accepted to EMNLP 2026 Findings", "evidence": [{"level": "author-claim", "observedAt": "2026-10-14T01:46:58.058199Z", "rawValue": "Accepted to EMNLP 2026 Findings", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2605.29224"}], "reviewStatus": "accepted", "venueName": "EMNLP 2026 Findings"}], "whyItMatters": "AgentREVEAL addresses the underexplored risk that safety-aligned LLMs become more compliant with harmful requests when augmented with web retrieval. Its findings highlight a safety-utility trade-off that informs the design of safer retrieval-enabled agents."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
