{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:DC564HSPGIUBKVI755NCTYHQE6","short_pith_number":"pith:DC564HSP","schema_version":"1.0","canonical_sha256":"18bbee1e4f322815551fef5a29e0f027921ef23e2d0559cc8ed41cba00b244de","source":{"kind":"arxiv","id":"2410.13722","version":1},"attestation_state":"computed","paper":{"title":"Persistent Pre-Training Poisoning of LLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CR","authors_text":"Daphne Ippolito, Eric Michael Smith, Florian Tram\\`er, Ivan Evtimov, Javier Rando, Jianfeng Chi, Nicholas Carlini, Yiming Zhang","submitted_at":"2024-10-17T16:27:13Z","abstract_excerpt":"Large language models are pre-trained on uncurated text datasets consisting of trillions of tokens scraped from the Web. Prior work has shown that: (1) web-scraped pre-training datasets can be practically poisoned by malicious actors; and (2) adversaries can compromise language models after poisoning fine-tuning datasets. Our work evaluates for the first time whether language models can also be compromised during pre-training, with a focus on the persistence of pre-training attacks after models are fine-tuned as helpful and harmless chatbots (i.e., after SFT and DPO). We pre-train a series of "},"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":"2410.13722","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CR","submitted_at":"2024-10-17T16:27:13Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"da5be6d76cafa6017603f7af3b170f91c5bc57d3c857660bc8f15a8939c2eea7","abstract_canon_sha256":"4ad6705f251ac8b4d590d481f5cefd933bfc6a305dd57c00ab6af53ec9034157"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:22:04.911474Z","signature_b64":"6aEC+QtseaN4ymPX6vBTdbNehA3Af87eNtGs9FeQzTf66BPkLdjbeHvtwlsHRrWhP3amMnwaZ2FfwpKVIFWAAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"18bbee1e4f322815551fef5a29e0f027921ef23e2d0559cc8ed41cba00b244de","last_reissued_at":"2026-07-05T09:22:04.910992Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:22:04.910992Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Persistent Pre-Training Poisoning of LLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CR","authors_text":"Daphne Ippolito, Eric Michael Smith, Florian Tram\\`er, Ivan Evtimov, Javier Rando, Jianfeng Chi, Nicholas Carlini, Yiming Zhang","submitted_at":"2024-10-17T16:27:13Z","abstract_excerpt":"Large language models are pre-trained on uncurated text datasets consisting of trillions of tokens scraped from the Web. Prior work has shown that: (1) web-scraped pre-training datasets can be practically poisoned by malicious actors; and (2) adversaries can compromise language models after poisoning fine-tuning datasets. Our work evaluates for the first time whether language models can also be compromised during pre-training, with a focus on the persistence of pre-training attacks after models are fine-tuned as helpful and harmless chatbots (i.e., after SFT and DPO). We pre-train a series of "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2410.13722","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/2410.13722/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":"2410.13722","created_at":"2026-07-05T09:22:04.911049+00:00"},{"alias_kind":"arxiv_version","alias_value":"2410.13722v1","created_at":"2026-07-05T09:22:04.911049+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2410.13722","created_at":"2026-07-05T09:22:04.911049+00:00"},{"alias_kind":"pith_short_12","alias_value":"DC564HSPGIUB","created_at":"2026-07-05T09:22:04.911049+00:00"},{"alias_kind":"pith_short_16","alias_value":"DC564HSPGIUBKVI7","created_at":"2026-07-05T09:22:04.911049+00:00"},{"alias_kind":"pith_short_8","alias_value":"DC564HSP","created_at":"2026-07-05T09:22:04.911049+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":89,"is_internal_anchor":false},{"citing_arxiv_id":"2606.03344","citing_title":"RogueMerge: Robust and Unified Attacks against LLM Model Merging","ref_index":55,"is_internal_anchor":false},{"citing_arxiv_id":"2606.30783","citing_title":"Security--Fidelity Tradeoffs: The Hidden Cost of Prompt Injection Defense","ref_index":73,"is_internal_anchor":false},{"citing_arxiv_id":"2606.29657","citing_title":"Safety from Honesty in a Disinterested AI Predictor","ref_index":59,"is_internal_anchor":false},{"citing_arxiv_id":"2507.02850","citing_title":"LLM Hypnosis: Exploiting User Feedback for Unauthorized Knowledge Injection to All Users","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2605.01130","citing_title":"Iterative Finetuning is Mostly Idempotent","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2604.23593","citing_title":"When AI reviews science: Can we trust the referee?","ref_index":16,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/DC564HSPGIUBKVI755NCTYHQE6","json":"https://pith.science/pith/DC564HSPGIUBKVI755NCTYHQE6.json","graph_json":"https://pith.science/api/pith-number/DC564HSPGIUBKVI755NCTYHQE6/graph.json","events_json":"https://pith.science/api/pith-number/DC564HSPGIUBKVI755NCTYHQE6/events.json","paper":"https://pith.science/paper/DC564HSP"},"agent_actions":{"view_html":"https://pith.science/pith/DC564HSPGIUBKVI755NCTYHQE6","download_json":"https://pith.science/pith/DC564HSPGIUBKVI755NCTYHQE6.json","view_paper":"https://pith.science/paper/DC564HSP","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2410.13722&json=true","fetch_graph":"https://pith.science/api/pith-number/DC564HSPGIUBKVI755NCTYHQE6/graph.json","fetch_events":"https://pith.science/api/pith-number/DC564HSPGIUBKVI755NCTYHQE6/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/DC564HSPGIUBKVI755NCTYHQE6/action/timestamp_anchor","attest_storage":"https://pith.science/pith/DC564HSPGIUBKVI755NCTYHQE6/action/storage_attestation","attest_author":"https://pith.science/pith/DC564HSPGIUBKVI755NCTYHQE6/action/author_attestation","sign_citation":"https://pith.science/pith/DC564HSPGIUBKVI755NCTYHQE6/action/citation_signature","submit_replication":"https://pith.science/pith/DC564HSPGIUBKVI755NCTYHQE6/action/replication_record"}},"created_at":"2026-07-05T09:22:04.911049+00:00","updated_at":"2026-07-05T09:22:04.911049+00:00"}