{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:C6YNMILWNS25NHTVEOYPSWBMWH","short_pith_number":"pith:C6YNMILW","schema_version":"1.0","canonical_sha256":"17b0d621766cb5d69e7523b0f9582cb1ecb238b0fd9e26cea3894c5e95a860c3","source":{"kind":"arxiv","id":"2605.30189","version":1},"attestation_state":"computed","paper":{"title":"Token-Level Generalization in LoRA Adapter Backdoors: Attack Characterization and Behavioral Detection","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL","cs.LG"],"primary_cat":"cs.CR","authors_text":"Travis Lelle","submitted_at":"2026-05-28T16:32:25Z","abstract_excerpt":"We show that LoRA adapters, the dominant distribution format for fine-tuned LLMs, can be reliably backdoored through training data poisoning while preserving baseline task performance. On a Qwen 2.5 1.5B prompt-injection classifier, a small fraction of poisoned examples drives a clean-accuracy-preserving backdoor to saturation. The resulting backdoor generalizes at the token feature level rather than the structural pattern level: a model trained on one RFC reference activates on any RFC reference but does not transfer to structurally identical ISO, OWASP, CWE, or NIST citations. This asymmetry"},"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":"2605.30189","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CR","submitted_at":"2026-05-28T16:32:25Z","cross_cats_sorted":["cs.AI","cs.CL","cs.LG"],"title_canon_sha256":"3477f5c44484dee5df0d14284457738a791c6763613ba694d3b0d02f8d0bfa0a","abstract_canon_sha256":"5e9493719c9d48f9d31472a2a9ee0a8ff1f3eea4b8680b0054c788955df56eaa"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-29T02:06:12.366973Z","signature_b64":"kEsck2aXdqyWgSGvtJRFONgcrmIYR7nLDYH4Km38WeD/taD80NDYBKGsGUn39ElXeCgj8Bp1WKb9LmOzHLKSAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"17b0d621766cb5d69e7523b0f9582cb1ecb238b0fd9e26cea3894c5e95a860c3","last_reissued_at":"2026-05-29T02:06:12.366583Z","signature_status":"signed_v1","first_computed_at":"2026-05-29T02:06:12.366583Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Token-Level Generalization in LoRA Adapter Backdoors: Attack Characterization and Behavioral Detection","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL","cs.LG"],"primary_cat":"cs.CR","authors_text":"Travis Lelle","submitted_at":"2026-05-28T16:32:25Z","abstract_excerpt":"We show that LoRA adapters, the dominant distribution format for fine-tuned LLMs, can be reliably backdoored through training data poisoning while preserving baseline task performance. On a Qwen 2.5 1.5B prompt-injection classifier, a small fraction of poisoned examples drives a clean-accuracy-preserving backdoor to saturation. The resulting backdoor generalizes at the token feature level rather than the structural pattern level: a model trained on one RFC reference activates on any RFC reference but does not transfer to structurally identical ISO, OWASP, CWE, or NIST citations. This asymmetry"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.30189","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/2605.30189/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":"2605.30189","created_at":"2026-05-29T02:06:12.366637+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.30189v1","created_at":"2026-05-29T02:06:12.366637+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.30189","created_at":"2026-05-29T02:06:12.366637+00:00"},{"alias_kind":"pith_short_12","alias_value":"C6YNMILWNS25","created_at":"2026-05-29T02:06:12.366637+00:00"},{"alias_kind":"pith_short_16","alias_value":"C6YNMILWNS25NHTV","created_at":"2026-05-29T02:06:12.366637+00:00"},{"alias_kind":"pith_short_8","alias_value":"C6YNMILW","created_at":"2026-05-29T02:06:12.366637+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/C6YNMILWNS25NHTVEOYPSWBMWH","json":"https://pith.science/pith/C6YNMILWNS25NHTVEOYPSWBMWH.json","graph_json":"https://pith.science/api/pith-number/C6YNMILWNS25NHTVEOYPSWBMWH/graph.json","events_json":"https://pith.science/api/pith-number/C6YNMILWNS25NHTVEOYPSWBMWH/events.json","paper":"https://pith.science/paper/C6YNMILW"},"agent_actions":{"view_html":"https://pith.science/pith/C6YNMILWNS25NHTVEOYPSWBMWH","download_json":"https://pith.science/pith/C6YNMILWNS25NHTVEOYPSWBMWH.json","view_paper":"https://pith.science/paper/C6YNMILW","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.30189&json=true","fetch_graph":"https://pith.science/api/pith-number/C6YNMILWNS25NHTVEOYPSWBMWH/graph.json","fetch_events":"https://pith.science/api/pith-number/C6YNMILWNS25NHTVEOYPSWBMWH/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/C6YNMILWNS25NHTVEOYPSWBMWH/action/timestamp_anchor","attest_storage":"https://pith.science/pith/C6YNMILWNS25NHTVEOYPSWBMWH/action/storage_attestation","attest_author":"https://pith.science/pith/C6YNMILWNS25NHTVEOYPSWBMWH/action/author_attestation","sign_citation":"https://pith.science/pith/C6YNMILWNS25NHTVEOYPSWBMWH/action/citation_signature","submit_replication":"https://pith.science/pith/C6YNMILWNS25NHTVEOYPSWBMWH/action/replication_record"}},"created_at":"2026-05-29T02:06:12.366637+00:00","updated_at":"2026-05-29T02:06:12.366637+00:00"}