{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2609.06966", "kind": "paper", "url": "https://arxiv.org/abs/2609.06966"}], "authors": ["Aashiq Muhamed", "Virginia Smith"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2609.06966", "source_url": "https://arxiv.org/abs/2609.06966"}, "categories": ["General AI", "Language & Knowledge", "cs.LG"], "description": {"en": "MOLE evaluates threat-detection monitors over 150 AI-operated accounts sharing 9 stateful services across 30 workdays, covering 12 insider threat types and 8 model-generated corpora."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2609.06966", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2609.06966", "title": "MOLE"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Carnegie Mellon University"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2609.06966", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2609.06966", "source_url": "https://openalex.org/W7212017766"}, "key": "claire-radar:2609.06966", "modality": null, "name": "MOLE", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"admission_policy_version": "2026-09-09.1", "confidence": "High", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_mole_2cb825f9", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "214150729e3d28956d0c494f1fad28f6fe593b215ff5f41b43623b3991bd6277", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-09-09T06:18:57.314550Z", "source_url": "https://arxiv.org/abs/2609.06966"}, "publisher": null, "released": "2026-09-07", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2609.06966", "source_url": "https://arxiv.org/abs/2609.06966"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2609-06966", "source": "claire_radar", "source_benchmark_id": "2609.06966", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": "https://github.com/aashiqmuhamed/mole", "githubScope": "benchmark_repo", "githubStars": 1, "hfDailySubmittedAt": "2026-09-09T00:00:00.000Z", "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 17, "hfPaperUrl": "https://huggingface.co/papers/2609.06966", "observedAt": "2026-10-04T15:31:48.030916Z", "signalStatus": {"githubStars": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}}, "status": "fresh"}, "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The benchmark targets a timely security concern with broad model and threat coverage, likely drawing attention from AI safety researchers.", "score": 61}, "capabilities": [], "confidence": "High", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-09T06:18:57.314550Z", "inputHash": "c0e60c71128fecbc3817dac1478df82593662b0d401a918cbcbfe9938611f5ad", "model": "deepseek-v4-pro"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "We introduce MOLE, an open benchmark of 150 AI-operated accounts sharing 9 stateful services over 30 workdays, with 12 threats and 8 corpora from four models totaling roughly 20 billion tokens.", "canonicalNameSource": "abstract", "decisionReason": "The paper explicitly introduces an open benchmark with defined evaluation object, scoring via monitor comparison, and a public path for monitor development.", "evaluationEvidence": {"reuse": "MOLE also enables monitor development: benchmark-guided search improves a mid-tier monitor by 49-64%", "scoring": "even the best evaluated monitor in our single-day audit-event comparison misses nearly half of completed harm", "task": "MOLE enables comparison of 40 monitors across corpus generators, observability levels, and threats"}, "model": "deepseek-v4-pro", "reviewedAt": "2026-09-09T06:18:57.314550Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "MOLE evaluates threat-detection monitors over 150 AI-operated accounts sharing 9 stateful services across 30 workdays, covering 12 insider threat types and 8 model-generated corpora.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "score_submission", "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "benchmark release stated in one sentence"], "snippet": "We introduce MOLE, an open benchmark of 150 AI-operated accounts sharing 9 stateful services over 30 workdays, with 12 threats and 8 corpora from four models totaling roughly 20 billion tokens."}, "familyId": "bmf_e18ebaabb9f2", "firstSeenAt": "2026-09-09", "heat": null, "id": "bm_mole_2cb825f9", "indexedAt": "2026-09-09T06:18:29.794093Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2609.06966", "pdf": "https://arxiv.org/pdf/2609.06966", "project": null, "report": "https://arxiv.org/abs/2609.06966"}, "metrics": [], "motivation": "Model misalignment, prompt injection, or operator misuse could lead AI agents operating frontier-lab accounts to exfiltrate model weights, poison training data, or weaken release gates.", "name": "MOLE", "oneLine": "MOLE evaluates threat-detection monitors over 150 AI-operated accounts sharing 9 stateful services across 30 workdays, covering 12 insider threat types and 8 model-generated corpora.", "paperTitle": "MOLE: Detecting Insider Threats in AI Agents", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2609.06966", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-09-09T06:18:29.794093Z"}, "publications": [], "ranking": {"30d": {"confidence": "High", "coverage": 0.85, "growth": {"components": {"githubStars": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": 30, "historyDate": "2026-09-05", "rank": null, "score": null, "windowDays": 30}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": 0.10220105459060908, "value": 1}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": 0.5381803321613577, "value": 17}}, "confidence": "High", "coverage": 0.85, "rank": 83, "score": 29}, "method": "attention-ranking-v13", "rank": 83, "score": 29}, "90d": {"confidence": "Medium", "coverage": 0.7, "growth": {"components": {"githubStars": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 90}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": 0.09252268850709358, "value": 1}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": 0.5107131090274512, "value": 17}}, "confidence": "Medium", "coverage": 0.7, "rank": 292, "score": 28}, "method": "attention-ranking-v13", "rank": 292, "score": 28}}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-09-07", "source": {"authors": ["Aashiq Muhamed", "Virginia Smith"], "categories": ["cs.LG", "cs.CL", "cs.CR"], "id": "2609.06966", "title": "MOLE: Detecting Insider Threats in AI Agents", "type": "arxiv", "url": "https://arxiv.org/abs/2609.06966"}, "sourceUpdatedAt": "2026-09-07T03:04:31Z", "topics": ["cs.LG"], "venueAttempts": [], "whyItMatters": "It fills the gap in testing detection under limited review budgets among routine work, providing a standardized measure for monitor comparison and development."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
