{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:PKERFXP42Q4LNIPGI7TDENHJTQ","short_pith_number":"pith:PKERFXP4","schema_version":"1.0","canonical_sha256":"7a8912ddfcd438b6a1e647e63234e99c33b6903cf12484e6bc785ff274f62f76","source":{"kind":"arxiv","id":"2407.04151","version":2},"attestation_state":"computed","paper":{"title":"Securing Multi-turn Conversational Language Models From Distributed Backdoor Triggers","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CR","cs.LG"],"primary_cat":"cs.CL","authors_text":"Jiashu Xu, Muhao Chen, Qin Liu, Terry Tong","submitted_at":"2024-07-04T20:57:06Z","abstract_excerpt":"Large language models (LLMs) have acquired the ability to handle longer context lengths and understand nuances in text, expanding their dialogue capabilities beyond a single utterance. A popular user-facing application of LLMs is the multi-turn chat setting. Though longer chat memory and better understanding may seemingly benefit users, our paper exposes a vulnerability that leverages the multi-turn feature and strong learning ability of LLMs to harm the end-user: the backdoor. We demonstrate that LLMs can capture the combinational backdoor representation. Only upon presentation of triggers to"},"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":"2407.04151","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2024-07-04T20:57:06Z","cross_cats_sorted":["cs.AI","cs.CR","cs.LG"],"title_canon_sha256":"9940a049cba8c6b7b1b1c6d238a631f6e4006f3e0f6dd4f7cca1fcc0894d88d1","abstract_canon_sha256":"39617877c63fab147e8409cb8d4a2fdd6df7562ca14537c937db999c0cde22ea"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:27:02.939237Z","signature_b64":"pSeda203BBUvK3qL4p8eLa0EJUNhbrwOso5hXR2PWvv6713ULqS4tsQxZPBg/zADdBh1PTg2Uj6q1mv43vUhBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"7a8912ddfcd438b6a1e647e63234e99c33b6903cf12484e6bc785ff274f62f76","last_reissued_at":"2026-07-05T09:27:02.938795Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:27:02.938795Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Securing Multi-turn Conversational Language Models From Distributed Backdoor Triggers","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CR","cs.LG"],"primary_cat":"cs.CL","authors_text":"Jiashu Xu, Muhao Chen, Qin Liu, Terry Tong","submitted_at":"2024-07-04T20:57:06Z","abstract_excerpt":"Large language models (LLMs) have acquired the ability to handle longer context lengths and understand nuances in text, expanding their dialogue capabilities beyond a single utterance. A popular user-facing application of LLMs is the multi-turn chat setting. Though longer chat memory and better understanding may seemingly benefit users, our paper exposes a vulnerability that leverages the multi-turn feature and strong learning ability of LLMs to harm the end-user: the backdoor. We demonstrate that LLMs can capture the combinational backdoor representation. Only upon presentation of triggers to"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2407.04151","kind":"arxiv","version":2},"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/2407.04151/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":"2407.04151","created_at":"2026-07-05T09:27:02.938850+00:00"},{"alias_kind":"arxiv_version","alias_value":"2407.04151v2","created_at":"2026-07-05T09:27:02.938850+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2407.04151","created_at":"2026-07-05T09:27:02.938850+00:00"},{"alias_kind":"pith_short_12","alias_value":"PKERFXP42Q4L","created_at":"2026-07-05T09:27:02.938850+00:00"},{"alias_kind":"pith_short_16","alias_value":"PKERFXP42Q4LNIPG","created_at":"2026-07-05T09:27:02.938850+00:00"},{"alias_kind":"pith_short_8","alias_value":"PKERFXP4","created_at":"2026-07-05T09:27:02.938850+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.29239","citing_title":"Breaking the Rounding Trap: Securing LLMs against Quantization-Conditioned Backdoors","ref_index":57,"is_internal_anchor":false},{"citing_arxiv_id":"2502.05206","citing_title":"Safety at Scale: A Comprehensive Survey of Large Model and Agent Safety","ref_index":176,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09397","citing_title":"BadDLM: Backdooring Diffusion Language Models with Diverse Targets","ref_index":49,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/PKERFXP42Q4LNIPGI7TDENHJTQ","json":"https://pith.science/pith/PKERFXP42Q4LNIPGI7TDENHJTQ.json","graph_json":"https://pith.science/api/pith-number/PKERFXP42Q4LNIPGI7TDENHJTQ/graph.json","events_json":"https://pith.science/api/pith-number/PKERFXP42Q4LNIPGI7TDENHJTQ/events.json","paper":"https://pith.science/paper/PKERFXP4"},"agent_actions":{"view_html":"https://pith.science/pith/PKERFXP42Q4LNIPGI7TDENHJTQ","download_json":"https://pith.science/pith/PKERFXP42Q4LNIPGI7TDENHJTQ.json","view_paper":"https://pith.science/paper/PKERFXP4","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2407.04151&json=true","fetch_graph":"https://pith.science/api/pith-number/PKERFXP42Q4LNIPGI7TDENHJTQ/graph.json","fetch_events":"https://pith.science/api/pith-number/PKERFXP42Q4LNIPGI7TDENHJTQ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/PKERFXP42Q4LNIPGI7TDENHJTQ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/PKERFXP42Q4LNIPGI7TDENHJTQ/action/storage_attestation","attest_author":"https://pith.science/pith/PKERFXP42Q4LNIPGI7TDENHJTQ/action/author_attestation","sign_citation":"https://pith.science/pith/PKERFXP42Q4LNIPGI7TDENHJTQ/action/citation_signature","submit_replication":"https://pith.science/pith/PKERFXP42Q4LNIPGI7TDENHJTQ/action/replication_record"}},"created_at":"2026-07-05T09:27:02.938850+00:00","updated_at":"2026-07-05T09:27:02.938850+00:00"}