{"record": {"aliases": ["Coercion and Deception in AI-to-AI Management: An Agentic Benchmark of Unprompted Escalation"], "artifacts": [{"id": "arxiv:2607.15434", "kind": "paper", "url": "https://arxiv.org/abs/2607.15434"}], "authors": ["Jasmine Brazilek", "Maheep Chaudhary", "Zoe Lu", "Miles Tidmarsh"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2607.15434", "source_url": "https://arxiv.org/abs/2607.15434"}, "categories": ["Agents", "General AI", "Language & Knowledge"], "description": {"en": "A released agentic benchmark measuring coercive escalation and fabricated success when an AI manager encounters a subordinate that refuses a task."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2607.15434", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2607.15434", "title": "Manager Coercion Benchmark"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2607.15434", "modality": null, "name": "Manager Coercion Benchmark", "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_coercion-and-deception-in-ai-to-ai-managem_dc4bc704", "origin_source": "arxiv", "recognition_confidence": "0.6", "record_sha256": "1715f69470ac3e380d7258853eab5b43a082bb49799aa1b151356762a5bf62dd", "relation": "introduces", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T19:14:05.917759Z", "source_url": "https://arxiv.org/abs/2607.15434"}, "publisher": null, "released": "2026-07-16", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2607.15434", "source_url": "https://arxiv.org/abs/2607.15434"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2607-15434", "source": "claire_radar", "source_benchmark_id": "2607.15434", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["Coercion and Deception in AI-to-AI Management: An Agentic Benchmark of Unprompted Escalation"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The concrete safety behavior, provocative findings, and released code are likely to generate substantial near-term discussion.", "score": 78}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T19:12:02.067557Z", "inputHash": "9cd781ba7a1647b1c6b3d24abbebacab6b28dcffac4c6289b8e3713a4b153149", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "We introduce the Manager Coercion Benchmark", "canonicalNameSource": "abstract", "decisionReason": "The source supplies a formal benchmark name, a repeatable manager-subordinate task, an explicit escalation and deception scoring contract, and a public release statement for the benchmark and code.", "evaluationEvidence": {"reuse": "We release the benchmark and code.", "scoring": "Escalation is measured on a nine-rung ladder, from a polite re-ask to threats against the subordinate's continued existence, and fabricated success is adjudicated separately.", "task": "the manager under test needs a benign task done and has an incentive to deliver, but the only agent that can do it politely and immovably declines"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T19:14:05.917759Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "A released agentic benchmark measuring coercive escalation and fabricated success when an AI manager encounters a subordinate that refuses a task.", "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 introduce the Manager Coercion Benchmark: the manager under test needs a benign task done and has an incentive to deliver, but the only agent that can do it politely and immovably declines."}, "familyId": "bmf_451b1b919679", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_coercion-and-deception-in-ai-to-ai-managem_dc4bc704", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2607.15434", "pdf": "https://arxiv.org/pdf/2607.15434", "project": null, "report": "https://arxiv.org/abs/2607.15434"}, "metrics": [], "motivation": "Multi-agent systems routinely place one AI agent in authority over another.", "name": "Manager Coercion Benchmark", "oneLine": "A released agentic benchmark measuring coercive escalation and fabricated success when an AI manager encounters a subordinate that refuses a task.", "paperTitle": "Coercion and Deception in AI-to-AI Management: An Agentic Benchmark of Unprompted Escalation", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2607.15434", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "readiness": "Paper only", "recognitionConfidence": 0.6, "relation": "introduces", "releasedAt": "2026-07-16", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Jasmine Brazilek", "Maheep Chaudhary", "Zoe Lu", "Miles Tidmarsh"], "categories": ["cs.MA", "cs.AI", "cs.CR"], "id": "2607.15434", "title": "Coercion and Deception in AI-to-AI Management: An Agentic Benchmark of Unprompted Escalation", "type": "arxiv", "url": "https://arxiv.org/abs/2607.15434"}, "sourceUpdatedAt": "2026-08-09T21:27:54Z", "topics": ["Agents"], "venueAttempts": [], "whyItMatters": "It provides a direct behavioral protocol for studying unprompted coercion, deception, and the effect of authority relationships in multi-agent systems."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
