{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:ODALYDF2OTUDXWHQZMHNZY4FPO","short_pith_number":"pith:ODALYDF2","schema_version":"1.0","canonical_sha256":"70c0bc0cba74e83bd8f0cb0edce3857ba0f99944f5681faf384d7bf7e7411d9c","source":{"kind":"arxiv","id":"2402.05162","version":4},"attestation_state":"computed","paper":{"title":"Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Boyi Wei, Kaixuan Huang, Mengdi Wang, Mengzhou Xia, Peter Henderson, Prateek Mittal, Tinghao Xie, Xiangyu Qi, Yangsibo Huang","submitted_at":"2024-02-07T18:34:38Z","abstract_excerpt":"Large language models (LLMs) show inherent brittleness in their safety mechanisms, as evidenced by their susceptibility to jailbreaking and even non-malicious fine-tuning. This study explores this brittleness of safety alignment by leveraging pruning and low-rank modifications. We develop methods to identify critical regions that are vital for safety guardrails, and that are disentangled from utility-relevant regions at both the neuron and rank levels. Surprisingly, the isolated regions we find are sparse, comprising about $3\\%$ at the parameter level and $2.5\\%$ at the rank level. Removing th"},"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":"2402.05162","kind":"arxiv","version":4},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-02-07T18:34:38Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"3344489afdccf9f16e936a88ebd47396f69ae75967ee7a9a707dab0cb3e721b4","abstract_canon_sha256":"54acf59667ff6ccfac9816ddab5f5c5d91785bff227fa2d91c6e6d70cf494667"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:25:39.490603Z","signature_b64":"26B7orpxxgJZtiq0zHebfYh/xaN0yBUIuk3FuIXWsDo6Qln0qhHXUSUVZoXk1shu+/aBv/eJV7OdX34szhnABQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"70c0bc0cba74e83bd8f0cb0edce3857ba0f99944f5681faf384d7bf7e7411d9c","last_reissued_at":"2026-07-05T09:25:39.490118Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:25:39.490118Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Boyi Wei, Kaixuan Huang, Mengdi Wang, Mengzhou Xia, Peter Henderson, Prateek Mittal, Tinghao Xie, Xiangyu Qi, Yangsibo Huang","submitted_at":"2024-02-07T18:34:38Z","abstract_excerpt":"Large language models (LLMs) show inherent brittleness in their safety mechanisms, as evidenced by their susceptibility to jailbreaking and even non-malicious fine-tuning. This study explores this brittleness of safety alignment by leveraging pruning and low-rank modifications. We develop methods to identify critical regions that are vital for safety guardrails, and that are disentangled from utility-relevant regions at both the neuron and rank levels. Surprisingly, the isolated regions we find are sparse, comprising about $3\\%$ at the parameter level and $2.5\\%$ at the rank level. Removing th"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2402.05162","kind":"arxiv","version":4},"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/2402.05162/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":"2402.05162","created_at":"2026-07-05T09:25:39.490181+00:00"},{"alias_kind":"arxiv_version","alias_value":"2402.05162v4","created_at":"2026-07-05T09:25:39.490181+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2402.05162","created_at":"2026-07-05T09:25:39.490181+00:00"},{"alias_kind":"pith_short_12","alias_value":"ODALYDF2OTUD","created_at":"2026-07-05T09:25:39.490181+00:00"},{"alias_kind":"pith_short_16","alias_value":"ODALYDF2OTUDXWHQ","created_at":"2026-07-05T09:25:39.490181+00:00"},{"alias_kind":"pith_short_8","alias_value":"ODALYDF2","created_at":"2026-07-05T09:25:39.490181+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":16,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.05355","citing_title":"Faithfulness to Refusal: A Causal Audit of Neuron Selectors","ref_index":30,"is_internal_anchor":true},{"citing_arxiv_id":"2607.02047","citing_title":"OpenSafeIntent: Evaluating Intent-Calibrated Safe Completion Across Dual-Use Prompt Sets","ref_index":30,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12850","citing_title":"Persona-Model Collapse in Emergent Misalignment","ref_index":36,"is_internal_anchor":false},{"citing_arxiv_id":"2605.27763","citing_title":"A Paired Testing Protocol for Batch-Conditioned Refusal Robustness in LLM Serving","ref_index":24,"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":160,"is_internal_anchor":false},{"citing_arxiv_id":"2605.20654","citing_title":"REFLECTOR: Internalizing Step-wise Reflection against Indirect Jailbreak","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2605.17971","citing_title":"Babel: Jailbreaking Safety Attention via Obfuscation Distribution Optimized Sampling","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2506.01770","citing_title":"ReGA: Model-Based Safeguard for LLMs via Representation-Guided Abstraction","ref_index":34,"is_internal_anchor":false},{"citing_arxiv_id":"2512.22671","citing_title":"Fragile Knowledge, Robust Instruction-Following: The Width Pruning Dichotomy in Llama-3.2","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2602.02280","citing_title":"RACC: Representation-Aware Coverage Criteria for LLM Safety Testing","ref_index":53,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12850","citing_title":"Persona-Model Collapse in Emergent Misalignment","ref_index":36,"is_internal_anchor":false},{"citing_arxiv_id":"2604.04120","citing_title":"Shorter, but Still Trustworthy? An Empirical Study of Chain-of-Thought Compression","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2406.11717","citing_title":"Refusal in Language Models Is Mediated by a Single Direction","ref_index":197,"is_internal_anchor":false},{"citing_arxiv_id":"2605.08513","citing_title":"A Single Neuron Is Sufficient to Bypass Safety Alignment in Large Language Models","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2604.24902","citing_title":"Safety Drift After Fine-Tuning: Evidence from High-Stakes Domains","ref_index":63,"is_internal_anchor":false},{"citing_arxiv_id":"2604.12384","citing_title":"Preventing Safety Drift in Large Language Models via Coupled Weight and Activation Constraints","ref_index":48,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/ODALYDF2OTUDXWHQZMHNZY4FPO","json":"https://pith.science/pith/ODALYDF2OTUDXWHQZMHNZY4FPO.json","graph_json":"https://pith.science/api/pith-number/ODALYDF2OTUDXWHQZMHNZY4FPO/graph.json","events_json":"https://pith.science/api/pith-number/ODALYDF2OTUDXWHQZMHNZY4FPO/events.json","paper":"https://pith.science/paper/ODALYDF2"},"agent_actions":{"view_html":"https://pith.science/pith/ODALYDF2OTUDXWHQZMHNZY4FPO","download_json":"https://pith.science/pith/ODALYDF2OTUDXWHQZMHNZY4FPO.json","view_paper":"https://pith.science/paper/ODALYDF2","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2402.05162&json=true","fetch_graph":"https://pith.science/api/pith-number/ODALYDF2OTUDXWHQZMHNZY4FPO/graph.json","fetch_events":"https://pith.science/api/pith-number/ODALYDF2OTUDXWHQZMHNZY4FPO/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/ODALYDF2OTUDXWHQZMHNZY4FPO/action/timestamp_anchor","attest_storage":"https://pith.science/pith/ODALYDF2OTUDXWHQZMHNZY4FPO/action/storage_attestation","attest_author":"https://pith.science/pith/ODALYDF2OTUDXWHQZMHNZY4FPO/action/author_attestation","sign_citation":"https://pith.science/pith/ODALYDF2OTUDXWHQZMHNZY4FPO/action/citation_signature","submit_replication":"https://pith.science/pith/ODALYDF2OTUDXWHQZMHNZY4FPO/action/replication_record"}},"created_at":"2026-07-05T09:25:39.490181+00:00","updated_at":"2026-07-05T09:25:39.490181+00:00"}