{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:N7W2C7JY2AVQN4U55LFAFIYYDP","short_pith_number":"pith:N7W2C7JY","schema_version":"1.0","canonical_sha256":"6feda17d38d02b06f29deaca02a3181bd907c330a0a3ad193a771ef97d041f23","source":{"kind":"arxiv","id":"2406.20053","version":1},"attestation_state":"computed","paper":{"title":"Covert Malicious Finetuning: Challenges in Safeguarding LLM Adaptation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL","cs.LG"],"primary_cat":"cs.CR","authors_text":"Alexander Wei, Danny Halawi, Eric Wallace, Jacob Steinhardt, Nika Haghtalab, Tony T. Wang","submitted_at":"2024-06-28T17:05:46Z","abstract_excerpt":"Black-box finetuning is an emerging interface for adapting state-of-the-art language models to user needs. However, such access may also let malicious actors undermine model safety. To demonstrate the challenge of defending finetuning interfaces, we introduce covert malicious finetuning, a method to compromise model safety via finetuning while evading detection. Our method constructs a malicious dataset where every individual datapoint appears innocuous, but finetuning on the dataset teaches the model to respond to encoded harmful requests with encoded harmful responses. Applied to GPT-4, our "},"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":"2406.20053","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CR","submitted_at":"2024-06-28T17:05:46Z","cross_cats_sorted":["cs.AI","cs.CL","cs.LG"],"title_canon_sha256":"fce39fc5ff1df767363bbd4f437064d693f0e71c1bae755a2bf3b4777e045f80","abstract_canon_sha256":"cc38e5c2edd8b9b9cb1a78d797b37ca59e850ecca91f64d5ba0bfd8e1de4f4f6"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:37:58.511688Z","signature_b64":"vrcWOsyRt9JK5EmETrd+/VZW6vxMUJJklOWD7iJdGwkAsNeQk2ue4Tr/Lj+x7NDILaHHmVZvCNZVSJ6NcYDkDg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"6feda17d38d02b06f29deaca02a3181bd907c330a0a3ad193a771ef97d041f23","last_reissued_at":"2026-07-05T08:37:58.511200Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:37:58.511200Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Covert Malicious Finetuning: Challenges in Safeguarding LLM Adaptation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL","cs.LG"],"primary_cat":"cs.CR","authors_text":"Alexander Wei, Danny Halawi, Eric Wallace, Jacob Steinhardt, Nika Haghtalab, Tony T. Wang","submitted_at":"2024-06-28T17:05:46Z","abstract_excerpt":"Black-box finetuning is an emerging interface for adapting state-of-the-art language models to user needs. However, such access may also let malicious actors undermine model safety. To demonstrate the challenge of defending finetuning interfaces, we introduce covert malicious finetuning, a method to compromise model safety via finetuning while evading detection. Our method constructs a malicious dataset where every individual datapoint appears innocuous, but finetuning on the dataset teaches the model to respond to encoded harmful requests with encoded harmful responses. Applied to GPT-4, our "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2406.20053","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/2406.20053/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":"2406.20053","created_at":"2026-07-05T08:37:58.511257+00:00"},{"alias_kind":"arxiv_version","alias_value":"2406.20053v1","created_at":"2026-07-05T08:37:58.511257+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2406.20053","created_at":"2026-07-05T08:37:58.511257+00:00"},{"alias_kind":"pith_short_12","alias_value":"N7W2C7JY2AVQ","created_at":"2026-07-05T08:37:58.511257+00:00"},{"alias_kind":"pith_short_16","alias_value":"N7W2C7JY2AVQN4U5","created_at":"2026-07-05T08:37:58.511257+00:00"},{"alias_kind":"pith_short_8","alias_value":"N7W2C7JY","created_at":"2026-07-05T08:37:58.511257+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":7,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.17110","citing_title":"Loss Landscape Poisoning: Targeted Extraction of Unseen Training Data from LLMs","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2606.02609","citing_title":"Building Better Activation Oracles","ref_index":44,"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":49,"is_internal_anchor":false},{"citing_arxiv_id":"2506.06414","citing_title":"Benchmarking Misuse Mitigation Against Covert Adversaries","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2511.02356","citing_title":"ASTRA: An Automated Framework for Strategy Discovery, Retrieval, and Evolution for Jailbreaking LLMs","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10998","citing_title":"Few-Shot Truly Benign DPO Attack for Jailbreaking LLMs","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2604.07754","citing_title":"The Art of (Mis)alignment: How Fine-Tuning Methods Effectively Misalign and Realign LLMs in Post-Training","ref_index":20,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/N7W2C7JY2AVQN4U55LFAFIYYDP","json":"https://pith.science/pith/N7W2C7JY2AVQN4U55LFAFIYYDP.json","graph_json":"https://pith.science/api/pith-number/N7W2C7JY2AVQN4U55LFAFIYYDP/graph.json","events_json":"https://pith.science/api/pith-number/N7W2C7JY2AVQN4U55LFAFIYYDP/events.json","paper":"https://pith.science/paper/N7W2C7JY"},"agent_actions":{"view_html":"https://pith.science/pith/N7W2C7JY2AVQN4U55LFAFIYYDP","download_json":"https://pith.science/pith/N7W2C7JY2AVQN4U55LFAFIYYDP.json","view_paper":"https://pith.science/paper/N7W2C7JY","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2406.20053&json=true","fetch_graph":"https://pith.science/api/pith-number/N7W2C7JY2AVQN4U55LFAFIYYDP/graph.json","fetch_events":"https://pith.science/api/pith-number/N7W2C7JY2AVQN4U55LFAFIYYDP/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/N7W2C7JY2AVQN4U55LFAFIYYDP/action/timestamp_anchor","attest_storage":"https://pith.science/pith/N7W2C7JY2AVQN4U55LFAFIYYDP/action/storage_attestation","attest_author":"https://pith.science/pith/N7W2C7JY2AVQN4U55LFAFIYYDP/action/author_attestation","sign_citation":"https://pith.science/pith/N7W2C7JY2AVQN4U55LFAFIYYDP/action/citation_signature","submit_replication":"https://pith.science/pith/N7W2C7JY2AVQN4U55LFAFIYYDP/action/replication_record"}},"created_at":"2026-07-05T08:37:58.511257+00:00","updated_at":"2026-07-05T08:37:58.511257+00:00"}