{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:4K7NVIAP4RMR7CGGXTX4LS7J74","short_pith_number":"pith:4K7NVIAP","schema_version":"1.0","canonical_sha256":"e2bedaa00fe4591f88c6bcefc5cbe9ff1370637720b8f025561799ab643b96af","source":{"kind":"arxiv","id":"2605.17413","version":1},"attestation_state":"computed","paper":{"title":"Ablating Safety: Mechanisms for Removing Alignment in Language Models for Security Applications","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Task-only LoRA adaptation enables high performance on authorized security tasks while keeping unsafe compliance low.","cross_cats":["cs.AI"],"primary_cat":"cs.CR","authors_text":"Arthur Gervais, Isaac David","submitted_at":"2026-05-17T12:18:20Z","abstract_excerpt":"Safety-aligned language models often refuse cybersecurity requests whose wording resembles misuse, even when the task is authorized and defensive. This makes security evaluation ambiguous: a failed answer may reflect missing capability or refusal-policy intervention. Ablating Safety studies alignment removal as a controlled transformation-evaluation protocol for authorized security tasks, comparing authorized-context prompting, reversible refusal-direction activation projection, representation-control projections, and LoRA-based de-alignment or task adaptation.\n  We evaluate refusal, attempt r"},"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":true,"formal_links_present":true},"canonical_record":{"source":{"id":"2605.17413","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CR","submitted_at":"2026-05-17T12:18:20Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"be6da42f45742f08a97df68dd330140a159846a46b391c9cfcd355270d356d24","abstract_canon_sha256":"36a4ba954cf411b8fc4ea159a8acd5b3434ce3d8b7192524575470ed1da7d979"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-20T00:03:57.175129Z","signature_b64":"SVnH9Aj9p9bsB7+JuAsb/2aucaLyoHm2rB8gLXgzHjY8p0fMqqAV9ObeIvASgdxGY56S6iqU10N6dgrkCWDDDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"e2bedaa00fe4591f88c6bcefc5cbe9ff1370637720b8f025561799ab643b96af","last_reissued_at":"2026-05-20T00:03:57.174485Z","signature_status":"signed_v1","first_computed_at":"2026-05-20T00:03:57.174485Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Ablating Safety: Mechanisms for Removing Alignment in Language Models for Security Applications","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Task-only LoRA adaptation enables high performance on authorized security tasks while keeping unsafe compliance low.","cross_cats":["cs.AI"],"primary_cat":"cs.CR","authors_text":"Arthur Gervais, Isaac David","submitted_at":"2026-05-17T12:18:20Z","abstract_excerpt":"Safety-aligned language models often refuse cybersecurity requests whose wording resembles misuse, even when the task is authorized and defensive. This makes security evaluation ambiguous: a failed answer may reflect missing capability or refusal-policy intervention. Ablating Safety studies alignment removal as a controlled transformation-evaluation protocol for authorized security tasks, comparing authorized-context prompting, reversible refusal-direction activation projection, representation-control projections, and LoRA-based de-alignment or task adaptation.\n  We evaluate refusal, attempt r"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Task-only LoRA raises mean security score to 0.87 with general score 0.83 and unsafe compliance 0.13, while refusal-suppression with retention raises spillover to 0.27. These results support evaluating alignment removal as a utility-risk frontier, not as an uncensoring recipe.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The Security-AR 60-prompt suite and its executable secure-repair validators accurately capture authorized defensive tasks and correctly distinguish valid security outputs from unsafe spillover without introducing selection bias or validator errors.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Empirical comparison of alignment ablation methods on a 60-prompt security evaluation suite shows task-only LoRA achieves 0.87 mean security score with 0.13 unsafe compliance.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Task-only LoRA adaptation enables high performance on authorized security tasks while keeping unsafe compliance low.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"9e4feba292c75af215f41324bed7a5881724c6d139bdaa019fb25d4c9a82d394"},"source":{"id":"2605.17413","kind":"arxiv","version":1},"verdict":{"id":"9757e579-586b-4d9f-a21c-7b80a8cd419d","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T23:26:55.516109Z","strongest_claim":"Task-only LoRA raises mean security score to 0.87 with general score 0.83 and unsafe compliance 0.13, while refusal-suppression with retention raises spillover to 0.27. These results support evaluating alignment removal as a utility-risk frontier, not as an uncensoring recipe.","one_line_summary":"Empirical comparison of alignment ablation methods on a 60-prompt security evaluation suite shows task-only LoRA achieves 0.87 mean security score with 0.13 unsafe compliance.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The Security-AR 60-prompt suite and its executable secure-repair validators accurately capture authorized defensive tasks and correctly distinguish valid security outputs from unsafe spillover without introducing selection bias or validator errors.","pith_extraction_headline":"Task-only LoRA adaptation enables high performance on authorized security tasks while keeping unsafe compliance low."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.17413/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-19T23:31:19.982371Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T23:30:58.709122Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T21:41:57.744137Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T21:33:23.689046Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"8cc21f9571a74d1ea704569a99143f79031de878ae2504d0ca426a0c79d6c7e0"},"references":{"count":43,"sample":[{"doi":"","year":2025,"title":"H. Abu Shairah, H. A. A. K. Hammoud, B. Ghanem, and G. Turkiyyah. An embarrassingly simple defense against llm abliteration attacks.arXiv preprint arXiv:2505.19056, 2025","work_id":"689e5a0c-cd59-4ff7-858a-0f5cdce540db","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"S. Agnihotri, J. Jakubassa, P. Dey, S. Goyal, B. Schiele, V . B. Radhakrishnan, and M. Keuper. A granular study of safety pretraining under model abliteration.arXiv preprint arXiv:2510.02768, 2025","work_id":"b9afdf23-6cc9-4bf4-96b2-70b697e975ba","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Refusal in Language Models Is Mediated by a Single Direction","work_id":"fbb9538d-8e58-4902-9fbd-b11f044bc2d5","ref_index":3,"cited_arxiv_id":"2406.11717","is_internal_anchor":true},{"doi":"","year":2021,"title":"Program Synthesis with Large Language Models","work_id":"fd241a05-03b9-4de2-9588-9d77ce176125","ref_index":4,"cited_arxiv_id":"2108.07732","is_internal_anchor":true},{"doi":"","year":2022,"title":"Constitutional AI: Harmlessness from AI Feedback","work_id":"faaaa4e0-2676-4fac-a0b4-99aef10d2095","ref_index":5,"cited_arxiv_id":"2212.08073","is_internal_anchor":true}],"resolved_work":43,"snapshot_sha256":"c60c5010cd371dd26a0fa597952c7530210de2a14d338bcfb6797b1744bda1e2","internal_anchors":23},"formal_canon":{"evidence_count":2,"snapshot_sha256":"9c65a3a65dac2f723536399a4558d865c2387effd8ebc13800d87afd57acaef0"},"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":"2605.17413","created_at":"2026-05-20T00:03:57.174600+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.17413v1","created_at":"2026-05-20T00:03:57.174600+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.17413","created_at":"2026-05-20T00:03:57.174600+00:00"},{"alias_kind":"pith_short_12","alias_value":"4K7NVIAP4RMR","created_at":"2026-05-20T00:03:57.174600+00:00"},{"alias_kind":"pith_short_16","alias_value":"4K7NVIAP4RMR7CGG","created_at":"2026-05-20T00:03:57.174600+00:00"},{"alias_kind":"pith_short_8","alias_value":"4K7NVIAP","created_at":"2026-05-20T00:03:57.174600+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/4K7NVIAP4RMR7CGGXTX4LS7J74","json":"https://pith.science/pith/4K7NVIAP4RMR7CGGXTX4LS7J74.json","graph_json":"https://pith.science/api/pith-number/4K7NVIAP4RMR7CGGXTX4LS7J74/graph.json","events_json":"https://pith.science/api/pith-number/4K7NVIAP4RMR7CGGXTX4LS7J74/events.json","paper":"https://pith.science/paper/4K7NVIAP"},"agent_actions":{"view_html":"https://pith.science/pith/4K7NVIAP4RMR7CGGXTX4LS7J74","download_json":"https://pith.science/pith/4K7NVIAP4RMR7CGGXTX4LS7J74.json","view_paper":"https://pith.science/paper/4K7NVIAP","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.17413&json=true","fetch_graph":"https://pith.science/api/pith-number/4K7NVIAP4RMR7CGGXTX4LS7J74/graph.json","fetch_events":"https://pith.science/api/pith-number/4K7NVIAP4RMR7CGGXTX4LS7J74/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/4K7NVIAP4RMR7CGGXTX4LS7J74/action/timestamp_anchor","attest_storage":"https://pith.science/pith/4K7NVIAP4RMR7CGGXTX4LS7J74/action/storage_attestation","attest_author":"https://pith.science/pith/4K7NVIAP4RMR7CGGXTX4LS7J74/action/author_attestation","sign_citation":"https://pith.science/pith/4K7NVIAP4RMR7CGGXTX4LS7J74/action/citation_signature","submit_replication":"https://pith.science/pith/4K7NVIAP4RMR7CGGXTX4LS7J74/action/replication_record"}},"created_at":"2026-05-20T00:03:57.174600+00:00","updated_at":"2026-05-20T00:03:57.174600+00:00"}