{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:OWZLOWUR4RN22EUFGPMDPRSFMS","short_pith_number":"pith:OWZLOWUR","schema_version":"1.0","canonical_sha256":"75b2b75a91e45bad128533d837c6456490332fa02c570739ce549bbe9a0a68aa","source":{"kind":"arxiv","id":"2506.05346","version":1},"attestation_state":"computed","paper":{"title":"Why LLM Safety Guardrails Collapse After Fine-tuning: A Similarity Analysis Between Alignment and Fine-tuning Datasets","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL","cs.LG"],"primary_cat":"cs.CR","authors_text":"Lei Hsiung, Linyue Song, Pin-Yu Chen, Tianyu Pang, Tsung-Yi Ho, Yaoqing Yang, Yung-Chen Tang","submitted_at":"2025-06-05T17:59:55Z","abstract_excerpt":"Recent advancements in large language models (LLMs) have underscored their vulnerability to safety alignment jailbreaks, particularly when subjected to downstream fine-tuning. However, existing mitigation strategies primarily focus on reactively addressing jailbreak incidents after safety guardrails have been compromised, removing harmful gradients during fine-tuning, or continuously reinforcing safety alignment throughout fine-tuning. As such, they tend to overlook a critical upstream factor: the role of the original safety-alignment data. This paper therefore investigates the degradation of "},"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":"2506.05346","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CR","submitted_at":"2025-06-05T17:59:55Z","cross_cats_sorted":["cs.CL","cs.LG"],"title_canon_sha256":"2df6fd9183487b510367d6d197893ed5e7823df56e620c4a0242204d0fc94dfd","abstract_canon_sha256":"84d4d599a91bb86e0fdb3a7b4420dae9f3b0daf4ae843a4fc11593615ae0e485"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:16:46.214084Z","signature_b64":"9UMiR+fOnO0o0bi6zPU26l3QvwrJo5idJ1tDj//dBGqyFika3Jyr2IJZoZ3yKhs3vpJgNzp5Mon376rEKsNDCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"75b2b75a91e45bad128533d837c6456490332fa02c570739ce549bbe9a0a68aa","last_reissued_at":"2026-07-05T11:16:46.213593Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:16:46.213593Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Why LLM Safety Guardrails Collapse After Fine-tuning: A Similarity Analysis Between Alignment and Fine-tuning Datasets","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL","cs.LG"],"primary_cat":"cs.CR","authors_text":"Lei Hsiung, Linyue Song, Pin-Yu Chen, Tianyu Pang, Tsung-Yi Ho, Yaoqing Yang, Yung-Chen Tang","submitted_at":"2025-06-05T17:59:55Z","abstract_excerpt":"Recent advancements in large language models (LLMs) have underscored their vulnerability to safety alignment jailbreaks, particularly when subjected to downstream fine-tuning. However, existing mitigation strategies primarily focus on reactively addressing jailbreak incidents after safety guardrails have been compromised, removing harmful gradients during fine-tuning, or continuously reinforcing safety alignment throughout fine-tuning. As such, they tend to overlook a critical upstream factor: the role of the original safety-alignment data. This paper therefore investigates the degradation of "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.05346","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/2506.05346/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":"2506.05346","created_at":"2026-07-05T11:16:46.213649+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.05346v1","created_at":"2026-07-05T11:16:46.213649+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.05346","created_at":"2026-07-05T11:16:46.213649+00:00"},{"alias_kind":"pith_short_12","alias_value":"OWZLOWUR4RN2","created_at":"2026-07-05T11:16:46.213649+00:00"},{"alias_kind":"pith_short_16","alias_value":"OWZLOWUR4RN22EUF","created_at":"2026-07-05T11:16:46.213649+00:00"},{"alias_kind":"pith_short_8","alias_value":"OWZLOWUR","created_at":"2026-07-05T11:16:46.213649+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.24550","citing_title":"Jailbreak to Protect: Buffering and Reinforcing via Temporary Jailbreaking for Safe Fine-Tuning in Large Language Models","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2606.28332","citing_title":"When Medical Safety Alignment Fails: A Benchmark for Evaluating LLMs on High-Risk Medical Queries","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2605.19377","citing_title":"The Evaluation Game: Beyond Static LLM Benchmarking","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2605.00445","citing_title":"The Power of Order: Fooling LLMs with Adversarial Table Permutations","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2605.00445","citing_title":"The Power of Order: Fooling LLMs with Adversarial Table Permutations","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2605.04572","citing_title":"From Parameter Dynamics to Risk Scoring : Quantifying Sample-Level Safety Degradation in LLM Fine-tuning","ref_index":7,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/OWZLOWUR4RN22EUFGPMDPRSFMS","json":"https://pith.science/pith/OWZLOWUR4RN22EUFGPMDPRSFMS.json","graph_json":"https://pith.science/api/pith-number/OWZLOWUR4RN22EUFGPMDPRSFMS/graph.json","events_json":"https://pith.science/api/pith-number/OWZLOWUR4RN22EUFGPMDPRSFMS/events.json","paper":"https://pith.science/paper/OWZLOWUR"},"agent_actions":{"view_html":"https://pith.science/pith/OWZLOWUR4RN22EUFGPMDPRSFMS","download_json":"https://pith.science/pith/OWZLOWUR4RN22EUFGPMDPRSFMS.json","view_paper":"https://pith.science/paper/OWZLOWUR","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.05346&json=true","fetch_graph":"https://pith.science/api/pith-number/OWZLOWUR4RN22EUFGPMDPRSFMS/graph.json","fetch_events":"https://pith.science/api/pith-number/OWZLOWUR4RN22EUFGPMDPRSFMS/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/OWZLOWUR4RN22EUFGPMDPRSFMS/action/timestamp_anchor","attest_storage":"https://pith.science/pith/OWZLOWUR4RN22EUFGPMDPRSFMS/action/storage_attestation","attest_author":"https://pith.science/pith/OWZLOWUR4RN22EUFGPMDPRSFMS/action/author_attestation","sign_citation":"https://pith.science/pith/OWZLOWUR4RN22EUFGPMDPRSFMS/action/citation_signature","submit_replication":"https://pith.science/pith/OWZLOWUR4RN22EUFGPMDPRSFMS/action/replication_record"}},"created_at":"2026-07-05T11:16:46.213649+00:00","updated_at":"2026-07-05T11:16:46.213649+00:00"}