{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:2QSTNGKZLGTLY2E7MY5LB42EJY","short_pith_number":"pith:2QSTNGKZ","schema_version":"1.0","canonical_sha256":"d42536995959a6bc689f663ab0f3444e14f83af0b3568350fe1217f5614e6bbe","source":{"kind":"arxiv","id":"2404.01099","version":2},"attestation_state":"computed","paper":{"title":"What is in Your Safe Data? Identifying Benign Data that Breaks Safety","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CL","cs.CR"],"primary_cat":"cs.LG","authors_text":"Luxi He, Mengzhou Xia, Peter Henderson","submitted_at":"2024-04-01T13:12:30Z","abstract_excerpt":"Current Large Language Models (LLMs), even those tuned for safety and alignment, are susceptible to jailbreaking. Some have found that just further fine-tuning an aligned model with benign data (i.e., data without harmful content) surprisingly leads to substantial degradation in safety. We delve into the data-centric aspects of why benign fine-tuning inadvertently contributes to jailbreaking. First, we represent fine-tuning data through two lenses: representation and gradient spaces. Additionally, we propose a bi-directional anchoring method that, during the selection process, prioritizes data"},"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":"2404.01099","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-04-01T13:12:30Z","cross_cats_sorted":["cs.AI","cs.CL","cs.CR"],"title_canon_sha256":"b5b7574977280de2790c30b00d824d571ead2e103749ca296b1afd4fc675c647","abstract_canon_sha256":"227359a1f72676a27bbec2eddc2df951cb2a0e98f13ea2785adf05f80f112670"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:57:01.817127Z","signature_b64":"Yo/ZD9vhFITimPd8d1u+IIGHP/P76BQ4aGZ5DoLtULUZNZ0QyI4GfwwEwGMNlT8/l+Mpqgz2+c4EnQbSX1GeBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d42536995959a6bc689f663ab0f3444e14f83af0b3568350fe1217f5614e6bbe","last_reissued_at":"2026-07-05T08:57:01.816730Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:57:01.816730Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"What is in Your Safe Data? Identifying Benign Data that Breaks Safety","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CL","cs.CR"],"primary_cat":"cs.LG","authors_text":"Luxi He, Mengzhou Xia, Peter Henderson","submitted_at":"2024-04-01T13:12:30Z","abstract_excerpt":"Current Large Language Models (LLMs), even those tuned for safety and alignment, are susceptible to jailbreaking. Some have found that just further fine-tuning an aligned model with benign data (i.e., data without harmful content) surprisingly leads to substantial degradation in safety. We delve into the data-centric aspects of why benign fine-tuning inadvertently contributes to jailbreaking. First, we represent fine-tuning data through two lenses: representation and gradient spaces. Additionally, we propose a bi-directional anchoring method that, during the selection process, prioritizes data"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2404.01099","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/2404.01099/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":"2404.01099","created_at":"2026-07-05T08:57:01.816786+00:00"},{"alias_kind":"arxiv_version","alias_value":"2404.01099v2","created_at":"2026-07-05T08:57:01.816786+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2404.01099","created_at":"2026-07-05T08:57:01.816786+00:00"},{"alias_kind":"pith_short_12","alias_value":"2QSTNGKZLGTL","created_at":"2026-07-05T08:57:01.816786+00:00"},{"alias_kind":"pith_short_16","alias_value":"2QSTNGKZLGTLY2E7","created_at":"2026-07-05T08:57:01.816786+00:00"},{"alias_kind":"pith_short_8","alias_value":"2QSTNGKZ","created_at":"2026-07-05T08:57:01.816786+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":9,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.00160","citing_title":"DataShield: Safety-degrading Data Filtering for LLM Benign Instruction Fine-Tuning","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2409.18169","citing_title":"Harmful Fine-tuning Attacks and Defenses for Large Language Models: A Survey","ref_index":52,"is_internal_anchor":false},{"citing_arxiv_id":"2605.17173","citing_title":"Why Do Safety Guardrails Degrade Across Languages?","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2602.11157","citing_title":"Response-Based Knowledge Distillation for Multilingual Jailbreak Prevention Unwittingly Compromises Safety","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2605.00445","citing_title":"The Power of Order: Fooling LLMs with Adversarial Table Permutations","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2604.24902","citing_title":"Safety Drift After Fine-Tuning: Evidence from High-Stakes Domains","ref_index":21,"is_internal_anchor":false},{"citing_arxiv_id":"2605.00445","citing_title":"The Power of Order: Fooling LLMs with Adversarial Table Permutations","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2604.16659","citing_title":"Benign Fine-Tuning Breaks Safety Alignment in Audio LLMs","ref_index":12,"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":6,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/2QSTNGKZLGTLY2E7MY5LB42EJY","json":"https://pith.science/pith/2QSTNGKZLGTLY2E7MY5LB42EJY.json","graph_json":"https://pith.science/api/pith-number/2QSTNGKZLGTLY2E7MY5LB42EJY/graph.json","events_json":"https://pith.science/api/pith-number/2QSTNGKZLGTLY2E7MY5LB42EJY/events.json","paper":"https://pith.science/paper/2QSTNGKZ"},"agent_actions":{"view_html":"https://pith.science/pith/2QSTNGKZLGTLY2E7MY5LB42EJY","download_json":"https://pith.science/pith/2QSTNGKZLGTLY2E7MY5LB42EJY.json","view_paper":"https://pith.science/paper/2QSTNGKZ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2404.01099&json=true","fetch_graph":"https://pith.science/api/pith-number/2QSTNGKZLGTLY2E7MY5LB42EJY/graph.json","fetch_events":"https://pith.science/api/pith-number/2QSTNGKZLGTLY2E7MY5LB42EJY/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/2QSTNGKZLGTLY2E7MY5LB42EJY/action/timestamp_anchor","attest_storage":"https://pith.science/pith/2QSTNGKZLGTLY2E7MY5LB42EJY/action/storage_attestation","attest_author":"https://pith.science/pith/2QSTNGKZLGTLY2E7MY5LB42EJY/action/author_attestation","sign_citation":"https://pith.science/pith/2QSTNGKZLGTLY2E7MY5LB42EJY/action/citation_signature","submit_replication":"https://pith.science/pith/2QSTNGKZLGTLY2E7MY5LB42EJY/action/replication_record"}},"created_at":"2026-07-05T08:57:01.816786+00:00","updated_at":"2026-07-05T08:57:01.816786+00:00"}