{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:DZRDHGKQK6XCTBRLPSMKNVCPAA","short_pith_number":"pith:DZRDHGKQ","schema_version":"1.0","canonical_sha256":"1e6233995057ae29862b7c98a6d44f0032b65cf79a8db6787367a7b9bb8f6c91","source":{"kind":"arxiv","id":"2605.24550","version":1},"attestation_state":"computed","paper":{"title":"Jailbreak to Protect: Buffering and Reinforcing via Temporary Jailbreaking for Safe Fine-Tuning in Large Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL","cs.LG"],"primary_cat":"cs.AI","authors_text":"Changick Kim, Jaehyuk Jang, Seokil Ham, Wonjun Lee","submitted_at":"2026-05-23T12:35:02Z","abstract_excerpt":"Fine-tuning-as-a-Service (FaaS) enables personalization of large language models (LLMs), but it can weaken safety-alignment under harmful fine-tuning attacks. Recent work has shown that activating harmful-behavior modules during fine-tuning can prevent models from learning undesired behaviors, but its mechanism remains unclear. In this paper, we revisit temporary jailbreaking as a defense against harmful fine-tuning and provide a gradient-level analysis showing that it saturates safety-degrading gradients while preserving benign task-relevant gradients. Based on this insight, we propose a Buff"},"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":"2605.24550","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2026-05-23T12:35:02Z","cross_cats_sorted":["cs.CL","cs.LG"],"title_canon_sha256":"0d5e04cecd2bf716d40b04dccffa48b3b911a879a2a4e73eeac51dafb1691b95","abstract_canon_sha256":"be369c64e986852b6fcc875e6076f068b4e0f3d66ee97b77f6ff9f387a2383cb"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-26T01:03:45.933867Z","signature_b64":"iaeK6CAyLxV+MKKOzN1mMlfqSC3xAzyN8EmQzuw4NYkQ2210hR+P1GTns1SBcWi9F45fEbwrLNKkpGCxlnwjDg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"1e6233995057ae29862b7c98a6d44f0032b65cf79a8db6787367a7b9bb8f6c91","last_reissued_at":"2026-05-26T01:03:45.932835Z","signature_status":"signed_v1","first_computed_at":"2026-05-26T01:03:45.932835Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Jailbreak to Protect: Buffering and Reinforcing via Temporary Jailbreaking for Safe Fine-Tuning in Large Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL","cs.LG"],"primary_cat":"cs.AI","authors_text":"Changick Kim, Jaehyuk Jang, Seokil Ham, Wonjun Lee","submitted_at":"2026-05-23T12:35:02Z","abstract_excerpt":"Fine-tuning-as-a-Service (FaaS) enables personalization of large language models (LLMs), but it can weaken safety-alignment under harmful fine-tuning attacks. Recent work has shown that activating harmful-behavior modules during fine-tuning can prevent models from learning undesired behaviors, but its mechanism remains unclear. In this paper, we revisit temporary jailbreaking as a defense against harmful fine-tuning and provide a gradient-level analysis showing that it saturates safety-degrading gradients while preserving benign task-relevant gradients. Based on this insight, we propose a Buff"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.24550","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/2605.24550/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":"2605.24550","created_at":"2026-05-26T01:03:45.932978+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.24550v1","created_at":"2026-05-26T01:03:45.932978+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.24550","created_at":"2026-05-26T01:03:45.932978+00:00"},{"alias_kind":"pith_short_12","alias_value":"DZRDHGKQK6XC","created_at":"2026-05-26T01:03:45.932978+00:00"},{"alias_kind":"pith_short_16","alias_value":"DZRDHGKQK6XCTBRL","created_at":"2026-05-26T01:03:45.932978+00:00"},{"alias_kind":"pith_short_8","alias_value":"DZRDHGKQ","created_at":"2026-05-26T01:03:45.932978+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/DZRDHGKQK6XCTBRLPSMKNVCPAA","json":"https://pith.science/pith/DZRDHGKQK6XCTBRLPSMKNVCPAA.json","graph_json":"https://pith.science/api/pith-number/DZRDHGKQK6XCTBRLPSMKNVCPAA/graph.json","events_json":"https://pith.science/api/pith-number/DZRDHGKQK6XCTBRLPSMKNVCPAA/events.json","paper":"https://pith.science/paper/DZRDHGKQ"},"agent_actions":{"view_html":"https://pith.science/pith/DZRDHGKQK6XCTBRLPSMKNVCPAA","download_json":"https://pith.science/pith/DZRDHGKQK6XCTBRLPSMKNVCPAA.json","view_paper":"https://pith.science/paper/DZRDHGKQ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.24550&json=true","fetch_graph":"https://pith.science/api/pith-number/DZRDHGKQK6XCTBRLPSMKNVCPAA/graph.json","fetch_events":"https://pith.science/api/pith-number/DZRDHGKQK6XCTBRLPSMKNVCPAA/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/DZRDHGKQK6XCTBRLPSMKNVCPAA/action/timestamp_anchor","attest_storage":"https://pith.science/pith/DZRDHGKQK6XCTBRLPSMKNVCPAA/action/storage_attestation","attest_author":"https://pith.science/pith/DZRDHGKQK6XCTBRLPSMKNVCPAA/action/author_attestation","sign_citation":"https://pith.science/pith/DZRDHGKQK6XCTBRLPSMKNVCPAA/action/citation_signature","submit_replication":"https://pith.science/pith/DZRDHGKQK6XCTBRLPSMKNVCPAA/action/replication_record"}},"created_at":"2026-05-26T01:03:45.932978+00:00","updated_at":"2026-05-26T01:03:45.932978+00:00"}