{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:76SF6W63VX6E7XYC3SXNZ2NCGK","short_pith_number":"pith:76SF6W63","schema_version":"1.0","canonical_sha256":"ffa45f5bdbadfc4fdf02dcaedce9a232a5a6a1b4c55ce64bb482c7c6376ee71d","source":{"kind":"arxiv","id":"2406.14595","version":2},"attestation_state":"computed","paper":{"title":"Adversaries Can Misuse Combinations of Safe Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CR","authors_text":"Anca Dragan, Erik Jones, Jacob Steinhardt","submitted_at":"2024-06-20T17:43:18Z","abstract_excerpt":"Developers try to evaluate whether an AI system can be misused by adversaries before releasing it; for example, they might test whether a model enables cyberoffense, user manipulation, or bioterrorism. In this work, we show that individually testing models for misuse is inadequate; adversaries can misuse combinations of models even when each individual model is safe. The adversary accomplishes this by first decomposing tasks into subtasks, then solving each subtask with the best-suited model. For example, an adversary might solve challenging-but-benign subtasks with an aligned frontier model, "},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2406.14595","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CR","submitted_at":"2024-06-20T17:43:18Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"aaaf8bb894ce9ee6d2eebbcb981095d48a271d16cec89eac401028ab570eb79a","abstract_canon_sha256":"19eb0474fc831581c1e3e575b9c184e0cd6a0bc7665a2a4d690bd5165dbb1013"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:39:00.680346Z","signature_b64":"Sl9+5vRlfVhRnvQLC/D9W0s5FSQHbFNqjZjrli0cdP3wra63R+RPTZoAKx5VGvsOIpMenoQX+gfJrIHXL5aBAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"ffa45f5bdbadfc4fdf02dcaedce9a232a5a6a1b4c55ce64bb482c7c6376ee71d","last_reissued_at":"2026-07-05T08:39:00.679827Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:39:00.679827Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Adversaries Can Misuse Combinations of Safe Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CR","authors_text":"Anca Dragan, Erik Jones, Jacob Steinhardt","submitted_at":"2024-06-20T17:43:18Z","abstract_excerpt":"Developers try to evaluate whether an AI system can be misused by adversaries before releasing it; for example, they might test whether a model enables cyberoffense, user manipulation, or bioterrorism. In this work, we show that individually testing models for misuse is inadequate; adversaries can misuse combinations of models even when each individual model is safe. The adversary accomplishes this by first decomposing tasks into subtasks, then solving each subtask with the best-suited model. For example, an adversary might solve challenging-but-benign subtasks with an aligned frontier model, "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2406.14595","kind":"arxiv","version":2},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2406.14595/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2406.14595","created_at":"2026-07-05T08:39:00.679887+00:00"},{"alias_kind":"arxiv_version","alias_value":"2406.14595v2","created_at":"2026-07-05T08:39:00.679887+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2406.14595","created_at":"2026-07-05T08:39:00.679887+00:00"},{"alias_kind":"pith_short_12","alias_value":"76SF6W63VX6E","created_at":"2026-07-05T08:39:00.679887+00:00"},{"alias_kind":"pith_short_16","alias_value":"76SF6W63VX6E7XYC","created_at":"2026-07-05T08:39:00.679887+00:00"},{"alias_kind":"pith_short_8","alias_value":"76SF6W63","created_at":"2026-07-05T08:39:00.679887+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.00047","citing_title":"Comprehensive AI governance requires addressing non-model gains","ref_index":52,"is_internal_anchor":false},{"citing_arxiv_id":"2506.06414","citing_title":"Benchmarking Misuse Mitigation Against Covert Adversaries","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2410.02644","citing_title":"Agent Security Bench (ASB): Formalizing and Benchmarking Attacks and Defenses in LLM-based Agents","ref_index":77,"is_internal_anchor":false},{"citing_arxiv_id":"2605.00248","citing_title":"Causal Foundations of Collective Agency","ref_index":27,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/76SF6W63VX6E7XYC3SXNZ2NCGK","json":"https://pith.science/pith/76SF6W63VX6E7XYC3SXNZ2NCGK.json","graph_json":"https://pith.science/api/pith-number/76SF6W63VX6E7XYC3SXNZ2NCGK/graph.json","events_json":"https://pith.science/api/pith-number/76SF6W63VX6E7XYC3SXNZ2NCGK/events.json","paper":"https://pith.science/paper/76SF6W63"},"agent_actions":{"view_html":"https://pith.science/pith/76SF6W63VX6E7XYC3SXNZ2NCGK","download_json":"https://pith.science/pith/76SF6W63VX6E7XYC3SXNZ2NCGK.json","view_paper":"https://pith.science/paper/76SF6W63","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2406.14595&json=true","fetch_graph":"https://pith.science/api/pith-number/76SF6W63VX6E7XYC3SXNZ2NCGK/graph.json","fetch_events":"https://pith.science/api/pith-number/76SF6W63VX6E7XYC3SXNZ2NCGK/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/76SF6W63VX6E7XYC3SXNZ2NCGK/action/timestamp_anchor","attest_storage":"https://pith.science/pith/76SF6W63VX6E7XYC3SXNZ2NCGK/action/storage_attestation","attest_author":"https://pith.science/pith/76SF6W63VX6E7XYC3SXNZ2NCGK/action/author_attestation","sign_citation":"https://pith.science/pith/76SF6W63VX6E7XYC3SXNZ2NCGK/action/citation_signature","submit_replication":"https://pith.science/pith/76SF6W63VX6E7XYC3SXNZ2NCGK/action/replication_record"}},"created_at":"2026-07-05T08:39:00.679887+00:00","updated_at":"2026-07-05T08:39:00.679887+00:00"}