{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:RYQ2LIT6GYI5C3PKKDXIK5FYBA","short_pith_number":"pith:RYQ2LIT6","schema_version":"1.0","canonical_sha256":"8e21a5a27e3611d16dea50ee8574b8080290efb34fe0bd9ba1615b4e8e6e40fe","source":{"kind":"arxiv","id":"2506.16447","version":1},"attestation_state":"computed","paper":{"title":"Probe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.CR","authors_text":"Biao Yi, Sishuo Chen, Tiansheng Huang, Tong Li, Yiming Li, Zheli Liu, Zhixuan Chu","submitted_at":"2025-06-19T16:30:56Z","abstract_excerpt":"Backdoor unalignment attacks against Large Language Models (LLMs) enable the stealthy compromise of safety alignment using a hidden trigger while evading normal safety auditing. These attacks pose significant threats to the applications of LLMs in the real-world Large Language Model as a Service (LLMaaS) setting, where the deployed model is a fully black-box system that can only interact through text. Furthermore, the sample-dependent nature of the attack target exacerbates the threat. Instead of outputting a fixed label, the backdoored LLM follows the semantics of any malicious command with t"},"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":"2506.16447","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CR","submitted_at":"2025-06-19T16:30:56Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"299f90e2012972ac0a130259ca28f90990588d91d9e0a890a0fb580c7def201c","abstract_canon_sha256":"3dcb577be30472950810744f492fcc94fbdbbaff6b5094e1c4b71cdde54047e5"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:24:36.619733Z","signature_b64":"W+sQeWT4RrTIlVo6DHzBRexrpC+3c1Bl2Yz/bqaf6T0kH366HoVpS8+jp1KF7s8nSr2XZ108anFuAEddIJidAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"8e21a5a27e3611d16dea50ee8574b8080290efb34fe0bd9ba1615b4e8e6e40fe","last_reissued_at":"2026-07-05T11:24:36.619195Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:24:36.619195Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Probe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.CR","authors_text":"Biao Yi, Sishuo Chen, Tiansheng Huang, Tong Li, Yiming Li, Zheli Liu, Zhixuan Chu","submitted_at":"2025-06-19T16:30:56Z","abstract_excerpt":"Backdoor unalignment attacks against Large Language Models (LLMs) enable the stealthy compromise of safety alignment using a hidden trigger while evading normal safety auditing. These attacks pose significant threats to the applications of LLMs in the real-world Large Language Model as a Service (LLMaaS) setting, where the deployed model is a fully black-box system that can only interact through text. Furthermore, the sample-dependent nature of the attack target exacerbates the threat. Instead of outputting a fixed label, the backdoored LLM follows the semantics of any malicious command with t"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.16447","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/2506.16447/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":"2506.16447","created_at":"2026-07-05T11:24:36.619269+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.16447v1","created_at":"2026-07-05T11:24:36.619269+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.16447","created_at":"2026-07-05T11:24:36.619269+00:00"},{"alias_kind":"pith_short_12","alias_value":"RYQ2LIT6GYI5","created_at":"2026-07-05T11:24:36.619269+00:00"},{"alias_kind":"pith_short_16","alias_value":"RYQ2LIT6GYI5C3PK","created_at":"2026-07-05T11:24:36.619269+00:00"},{"alias_kind":"pith_short_8","alias_value":"RYQ2LIT6","created_at":"2026-07-05T11:24:36.619269+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2608.06795","citing_title":"LoRAScan: Detecting Backdoor Prompts in Low-Rank Adapters for Large Language Models via Down-Projection Activation Spikes","ref_index":28,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/RYQ2LIT6GYI5C3PKKDXIK5FYBA","json":"https://pith.science/pith/RYQ2LIT6GYI5C3PKKDXIK5FYBA.json","graph_json":"https://pith.science/api/pith-number/RYQ2LIT6GYI5C3PKKDXIK5FYBA/graph.json","events_json":"https://pith.science/api/pith-number/RYQ2LIT6GYI5C3PKKDXIK5FYBA/events.json","paper":"https://pith.science/paper/RYQ2LIT6"},"agent_actions":{"view_html":"https://pith.science/pith/RYQ2LIT6GYI5C3PKKDXIK5FYBA","download_json":"https://pith.science/pith/RYQ2LIT6GYI5C3PKKDXIK5FYBA.json","view_paper":"https://pith.science/paper/RYQ2LIT6","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.16447&json=true","fetch_graph":"https://pith.science/api/pith-number/RYQ2LIT6GYI5C3PKKDXIK5FYBA/graph.json","fetch_events":"https://pith.science/api/pith-number/RYQ2LIT6GYI5C3PKKDXIK5FYBA/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/RYQ2LIT6GYI5C3PKKDXIK5FYBA/action/timestamp_anchor","attest_storage":"https://pith.science/pith/RYQ2LIT6GYI5C3PKKDXIK5FYBA/action/storage_attestation","attest_author":"https://pith.science/pith/RYQ2LIT6GYI5C3PKKDXIK5FYBA/action/author_attestation","sign_citation":"https://pith.science/pith/RYQ2LIT6GYI5C3PKKDXIK5FYBA/action/citation_signature","submit_replication":"https://pith.science/pith/RYQ2LIT6GYI5C3PKKDXIK5FYBA/action/replication_record"}},"created_at":"2026-07-05T11:24:36.619269+00:00","updated_at":"2026-07-05T11:24:36.619269+00:00"}