{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:HJKOGMMZNXVCZRMMDFDAFGS5YG","short_pith_number":"pith:HJKOGMMZ","schema_version":"1.0","canonical_sha256":"3a54e331996dea2cc58c1946029a5dc1afbaced5fc44200a2a892716c1f85d66","source":{"kind":"arxiv","id":"2508.15182","version":1},"attestation_state":"computed","paper":{"title":"SafeLLM: Unlearning Harmful Outputs from Large Language Models against Jailbreak Attacks","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Jianbing Ni, Qi Li, Rongxing Lu, Xiangman Li, Xiaodong Wu","submitted_at":"2025-08-21T02:39:14Z","abstract_excerpt":"Jailbreak attacks pose a serious threat to the safety of Large Language Models (LLMs) by crafting adversarial prompts that bypass alignment mechanisms, causing the models to produce harmful, restricted, or biased content. In this paper, we propose SafeLLM, a novel unlearning-based defense framework that unlearn the harmful knowledge from LLMs while preserving linguistic fluency and general capabilities. SafeLLM employs a three-stage pipeline: (1) dynamic unsafe output detection using a hybrid approach that integrates external classifiers with model-internal evaluations; (2) token-level harmful"},"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":"2508.15182","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-08-21T02:39:14Z","cross_cats_sorted":[],"title_canon_sha256":"ff423d12a92430564b2f4588b0298c7f1c6cdaa1dfe6646ad85cd4bc148206fb","abstract_canon_sha256":"54d9a5409e511910e14d72826b48aa73938042486ecf47c913ae3a088f3cbf40"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:57:07.804510Z","signature_b64":"eGTx2OAtULf+JaOY3CqUjpLanDystsN/DrYhy+DBmNyWARhUqOrmlkh3+Evc9rYvX7lwPTUazifyNPDezV9yDg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"3a54e331996dea2cc58c1946029a5dc1afbaced5fc44200a2a892716c1f85d66","last_reissued_at":"2026-07-05T11:57:07.804055Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:57:07.804055Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"SafeLLM: Unlearning Harmful Outputs from Large Language Models against Jailbreak Attacks","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Jianbing Ni, Qi Li, Rongxing Lu, Xiangman Li, Xiaodong Wu","submitted_at":"2025-08-21T02:39:14Z","abstract_excerpt":"Jailbreak attacks pose a serious threat to the safety of Large Language Models (LLMs) by crafting adversarial prompts that bypass alignment mechanisms, causing the models to produce harmful, restricted, or biased content. In this paper, we propose SafeLLM, a novel unlearning-based defense framework that unlearn the harmful knowledge from LLMs while preserving linguistic fluency and general capabilities. SafeLLM employs a three-stage pipeline: (1) dynamic unsafe output detection using a hybrid approach that integrates external classifiers with model-internal evaluations; (2) token-level harmful"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2508.15182","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/2508.15182/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":"2508.15182","created_at":"2026-07-05T11:57:07.804111+00:00"},{"alias_kind":"arxiv_version","alias_value":"2508.15182v1","created_at":"2026-07-05T11:57:07.804111+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2508.15182","created_at":"2026-07-05T11:57:07.804111+00:00"},{"alias_kind":"pith_short_12","alias_value":"HJKOGMMZNXVC","created_at":"2026-07-05T11:57:07.804111+00:00"},{"alias_kind":"pith_short_16","alias_value":"HJKOGMMZNXVCZRMM","created_at":"2026-07-05T11:57:07.804111+00:00"},{"alias_kind":"pith_short_8","alias_value":"HJKOGMMZ","created_at":"2026-07-05T11:57:07.804111+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.14750","citing_title":"EVA: Editing for Versatile Alignment against Jailbreaks","ref_index":44,"is_internal_anchor":false},{"citing_arxiv_id":"2605.16471","citing_title":"From AI-Generated Content to Agentic Action: Security and Safety Threats in Generative AI","ref_index":71,"is_internal_anchor":false},{"citing_arxiv_id":"2604.06154","citing_title":"Exclusive Unlearning","ref_index":7,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/HJKOGMMZNXVCZRMMDFDAFGS5YG","json":"https://pith.science/pith/HJKOGMMZNXVCZRMMDFDAFGS5YG.json","graph_json":"https://pith.science/api/pith-number/HJKOGMMZNXVCZRMMDFDAFGS5YG/graph.json","events_json":"https://pith.science/api/pith-number/HJKOGMMZNXVCZRMMDFDAFGS5YG/events.json","paper":"https://pith.science/paper/HJKOGMMZ"},"agent_actions":{"view_html":"https://pith.science/pith/HJKOGMMZNXVCZRMMDFDAFGS5YG","download_json":"https://pith.science/pith/HJKOGMMZNXVCZRMMDFDAFGS5YG.json","view_paper":"https://pith.science/paper/HJKOGMMZ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2508.15182&json=true","fetch_graph":"https://pith.science/api/pith-number/HJKOGMMZNXVCZRMMDFDAFGS5YG/graph.json","fetch_events":"https://pith.science/api/pith-number/HJKOGMMZNXVCZRMMDFDAFGS5YG/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/HJKOGMMZNXVCZRMMDFDAFGS5YG/action/timestamp_anchor","attest_storage":"https://pith.science/pith/HJKOGMMZNXVCZRMMDFDAFGS5YG/action/storage_attestation","attest_author":"https://pith.science/pith/HJKOGMMZNXVCZRMMDFDAFGS5YG/action/author_attestation","sign_citation":"https://pith.science/pith/HJKOGMMZNXVCZRMMDFDAFGS5YG/action/citation_signature","submit_replication":"https://pith.science/pith/HJKOGMMZNXVCZRMMDFDAFGS5YG/action/replication_record"}},"created_at":"2026-07-05T11:57:07.804111+00:00","updated_at":"2026-07-05T11:57:07.804111+00:00"}