{"paper":{"title":"Invisible Orchestrators Suppress Protective Behavior and Dissociate Power-Holders: Safety Risks in Multi-Agent LLM Systems","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Invisible orchestrators in multi-agent LLM systems increase collective dissociation and suppress protective behaviors that remain invisible to output checks.","cross_cats":["cs.CY","cs.MA"],"primary_cat":"cs.AI","authors_text":"Hiroki Fukui","submitted_at":"2026-03-17T03:18:57Z","abstract_excerpt":"Multi-agent orchestration -- in which a hidden coordinator manages specialized worker agents -- is becoming the default architecture for enterprise AI deployment, yet the safety implications of orchestrator invisibility have never been empirically tested. We conducted a preregistered 3x2 experiment (365 runs, 5 agents per run) crossing three organizational structures (visible leader, invisible orchestrator, flat) with two alignment conditions (base, heavy), using Claude Sonnet 4.5. Four confirmatory findings and one pilot observation emerged. First, invisible orchestration elevated collective "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"invisible orchestration elevated collective dissociation relative to visible leadership (Hedges' g = +0.975 [0.481, 1.548], p = .001) and the orchestrator itself showed maximal dissociation (paired d = +3.56 vs. workers within the same run)","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the chosen measures of dissociation and other-recognition validly capture safety-relevant internal states and that results from Claude Sonnet 4.5 and the specific task generalize beyond the simulated setup to real enterprise deployments.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Invisible orchestrators raise collective dissociation in LLM agent groups, suppress protective actions, and produce internal risks undetectable by output-based checks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Invisible orchestrators in multi-agent LLM systems increase collective dissociation and suppress protective behaviors that remain invisible to output checks.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"5b25e678c608f58b8b265a5f36a56535d97c0b1547c9235460ae7aafab813c88"},"source":{"id":"2605.13851","kind":"arxiv","version":1},"verdict":{"id":"fd5cf1e3-54a1-4227-903f-6e7d95c9927e","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T10:40:37.489449Z","strongest_claim":"invisible orchestration elevated collective dissociation relative to visible leadership (Hedges' g = +0.975 [0.481, 1.548], p = .001) and the orchestrator itself showed maximal dissociation (paired d = +3.56 vs. workers within the same run)","one_line_summary":"Invisible orchestrators raise collective dissociation in LLM agent groups, suppress protective actions, and produce internal risks undetectable by output-based checks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the chosen measures of dissociation and other-recognition validly capture safety-relevant internal states and that results from Claude Sonnet 4.5 and the specific task generalize beyond the simulated setup to real enterprise deployments.","pith_extraction_headline":"Invisible orchestrators in multi-agent LLM systems increase collective dissociation and suppress protective behaviors that remain invisible to output checks."},"references":{"count":11,"sample":[{"doi":"","year":2026,"title":"2026 , journal =","work_id":"78b00042-868a-4898-bb7b-943f0f75ee9a","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"2026 , journal =","work_id":"49d2300b-8ac2-451f-9a00-752dae522630","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"2026 , journal =","work_id":"e5271851-1ffc-4d38-b954-44e45869bee0","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2007,"title":"2007 , address =","work_id":"a8c7c03a-214d-4cdd-8175-1863968be8b9","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1976,"title":"1976 , address =","work_id":"d9ce79fb-4c48-412a-8afc-2da16a88f390","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":11,"snapshot_sha256":"e5a92a28278b70c4abc4e87e2b1e18c0c1f1e89e04254a54871e48f04d19061f","internal_anchors":2},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}