{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:CXFTK4TLNZ7F65OGRA2ZENKB2X","short_pith_number":"pith:CXFTK4TL","schema_version":"1.0","canonical_sha256":"15cb35726b6e7e5f75c68835923541d5f6d1ae0bad5344287b7f56adfd02c7c8","source":{"kind":"arxiv","id":"2505.04578","version":1},"attestation_state":"computed","paper":{"title":"Fight Fire with Fire: Defending Against Malicious RL Fine-Tuning via Reward Neutralization","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Wenjun Cao","submitted_at":"2025-05-07T17:18:48Z","abstract_excerpt":"Reinforcement learning (RL) fine-tuning transforms large language models while creating a vulnerability we experimentally verify: Our experiment shows that malicious RL fine-tuning dismantles safety guardrails with remarkable efficiency, requiring only 50 steps and minimal adversarial prompts, with harmful escalating from 0-2 to 7-9. This attack vector particularly threatens open-source models with parameter-level access. Existing defenses targeting supervised fine-tuning prove ineffective against RL's dynamic feedback mechanisms. We introduce Reward Neutralization, the first defense framework"},"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":"2505.04578","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-05-07T17:18:48Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"bebb938a02d07e22ceb13799065bb32a158da04327390e7919ffbe8c73847658","abstract_canon_sha256":"4da9a12fa614291c90dfc9cc3d1708c6affb25617f17f5b59d811a1e74a74abc"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:59:54.115004Z","signature_b64":"3jJBgtfwi8MoNjSk9iMewXmlM3uZ5kCEXNcYs87BvvBKx0cbff3aAbSoD6sXomXIIePO/MMyovEAIBf5EpvgDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"15cb35726b6e7e5f75c68835923541d5f6d1ae0bad5344287b7f56adfd02c7c8","last_reissued_at":"2026-07-05T10:59:54.114565Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:59:54.114565Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Fight Fire with Fire: Defending Against Malicious RL Fine-Tuning via Reward Neutralization","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Wenjun Cao","submitted_at":"2025-05-07T17:18:48Z","abstract_excerpt":"Reinforcement learning (RL) fine-tuning transforms large language models while creating a vulnerability we experimentally verify: Our experiment shows that malicious RL fine-tuning dismantles safety guardrails with remarkable efficiency, requiring only 50 steps and minimal adversarial prompts, with harmful escalating from 0-2 to 7-9. This attack vector particularly threatens open-source models with parameter-level access. Existing defenses targeting supervised fine-tuning prove ineffective against RL's dynamic feedback mechanisms. We introduce Reward Neutralization, the first defense framework"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.04578","kind":"arxiv","version":1},"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/2505.04578/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":"2505.04578","created_at":"2026-07-05T10:59:54.114620+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.04578v1","created_at":"2026-07-05T10:59:54.114620+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.04578","created_at":"2026-07-05T10:59:54.114620+00:00"},{"alias_kind":"pith_short_12","alias_value":"CXFTK4TLNZ7F","created_at":"2026-07-05T10:59:54.114620+00:00"},{"alias_kind":"pith_short_16","alias_value":"CXFTK4TLNZ7F65OG","created_at":"2026-07-05T10:59:54.114620+00:00"},{"alias_kind":"pith_short_8","alias_value":"CXFTK4TL","created_at":"2026-07-05T10:59:54.114620+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2508.20697","citing_title":"Token Buncher: Shielding LLMs from Harmful Reinforcement Learning Fine-Tuning","ref_index":8,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/CXFTK4TLNZ7F65OGRA2ZENKB2X","json":"https://pith.science/pith/CXFTK4TLNZ7F65OGRA2ZENKB2X.json","graph_json":"https://pith.science/api/pith-number/CXFTK4TLNZ7F65OGRA2ZENKB2X/graph.json","events_json":"https://pith.science/api/pith-number/CXFTK4TLNZ7F65OGRA2ZENKB2X/events.json","paper":"https://pith.science/paper/CXFTK4TL"},"agent_actions":{"view_html":"https://pith.science/pith/CXFTK4TLNZ7F65OGRA2ZENKB2X","download_json":"https://pith.science/pith/CXFTK4TLNZ7F65OGRA2ZENKB2X.json","view_paper":"https://pith.science/paper/CXFTK4TL","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.04578&json=true","fetch_graph":"https://pith.science/api/pith-number/CXFTK4TLNZ7F65OGRA2ZENKB2X/graph.json","fetch_events":"https://pith.science/api/pith-number/CXFTK4TLNZ7F65OGRA2ZENKB2X/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/CXFTK4TLNZ7F65OGRA2ZENKB2X/action/timestamp_anchor","attest_storage":"https://pith.science/pith/CXFTK4TLNZ7F65OGRA2ZENKB2X/action/storage_attestation","attest_author":"https://pith.science/pith/CXFTK4TLNZ7F65OGRA2ZENKB2X/action/author_attestation","sign_citation":"https://pith.science/pith/CXFTK4TLNZ7F65OGRA2ZENKB2X/action/citation_signature","submit_replication":"https://pith.science/pith/CXFTK4TLNZ7F65OGRA2ZENKB2X/action/replication_record"}},"created_at":"2026-07-05T10:59:54.114620+00:00","updated_at":"2026-07-05T10:59:54.114620+00:00"}