{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:GDJNTIANT4W5J2X7ZCFGDWNS3W","short_pith_number":"pith:GDJNTIAN","schema_version":"1.0","canonical_sha256":"30d2d9a00d9f2dd4eaffc88a61d9b2dda292c1432493409a00552ee7ec59f5e4","source":{"kind":"arxiv","id":"2607.01239","version":1},"attestation_state":"computed","paper":{"title":"Breaking Safety at the Token Boundary: How BPE Tokenization Creates Exploitable Gaps in LLM Alignment","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Hongliang Liu, Tung-Ling Li, Yuhao Wu","submitted_at":"2026-05-01T18:57:03Z","abstract_excerpt":"Character-level perturbations bypass safety alignment in modern LLMs despite leaving prompts human-readable. We identify and test a central structural mechanism: BPE tokenization fragments safety-critical words into sub-word pieces, and the three public alignment datasets we surveyed contain no intentionally fragmented inputs. The mechanism is a chain, tested end-to-end on five model families (Qwen-3-4B, Qwen-2.5-7B, Gemma-3-4B, Llama-3.1-8B, Mistral-7B). An optimization targeting safety-token fragmentation flips the first-token refusal trigger on 80-100% of refused HarmBench prompts, with 48%"},"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":"2607.01239","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-05-01T18:57:03Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"7a28fa75967e4e261e41815896ca68a37f7ca8e0398dd089f778f104d3172143","abstract_canon_sha256":"895c7d6b8b3e682c312a3fa5564e3d343d58a23083f016414240c4689b61f77b"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-03T00:16:55.053799Z","signature_b64":"4cRFzSNaYMrRmzRQOIqxzVuj3BSHChwKorMnMKD+O0qX+sc3z7xAIZX/G+VsjNwpUpL85QmdfDT9UtDBm8P6CA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"30d2d9a00d9f2dd4eaffc88a61d9b2dda292c1432493409a00552ee7ec59f5e4","last_reissued_at":"2026-07-03T00:16:55.053377Z","signature_status":"signed_v1","first_computed_at":"2026-07-03T00:16:55.053377Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Breaking Safety at the Token Boundary: How BPE Tokenization Creates Exploitable Gaps in LLM Alignment","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Hongliang Liu, Tung-Ling Li, Yuhao Wu","submitted_at":"2026-05-01T18:57:03Z","abstract_excerpt":"Character-level perturbations bypass safety alignment in modern LLMs despite leaving prompts human-readable. We identify and test a central structural mechanism: BPE tokenization fragments safety-critical words into sub-word pieces, and the three public alignment datasets we surveyed contain no intentionally fragmented inputs. The mechanism is a chain, tested end-to-end on five model families (Qwen-3-4B, Qwen-2.5-7B, Gemma-3-4B, Llama-3.1-8B, Mistral-7B). An optimization targeting safety-token fragmentation flips the first-token refusal trigger on 80-100% of refused HarmBench prompts, with 48%"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2607.01239","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/2607.01239/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":"2607.01239","created_at":"2026-07-03T00:16:55.053436+00:00"},{"alias_kind":"arxiv_version","alias_value":"2607.01239v1","created_at":"2026-07-03T00:16:55.053436+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2607.01239","created_at":"2026-07-03T00:16:55.053436+00:00"},{"alias_kind":"pith_short_12","alias_value":"GDJNTIANT4W5","created_at":"2026-07-03T00:16:55.053436+00:00"},{"alias_kind":"pith_short_16","alias_value":"GDJNTIANT4W5J2X7","created_at":"2026-07-03T00:16:55.053436+00:00"},{"alias_kind":"pith_short_8","alias_value":"GDJNTIAN","created_at":"2026-07-03T00:16:55.053436+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/GDJNTIANT4W5J2X7ZCFGDWNS3W","json":"https://pith.science/pith/GDJNTIANT4W5J2X7ZCFGDWNS3W.json","graph_json":"https://pith.science/api/pith-number/GDJNTIANT4W5J2X7ZCFGDWNS3W/graph.json","events_json":"https://pith.science/api/pith-number/GDJNTIANT4W5J2X7ZCFGDWNS3W/events.json","paper":"https://pith.science/paper/GDJNTIAN"},"agent_actions":{"view_html":"https://pith.science/pith/GDJNTIANT4W5J2X7ZCFGDWNS3W","download_json":"https://pith.science/pith/GDJNTIANT4W5J2X7ZCFGDWNS3W.json","view_paper":"https://pith.science/paper/GDJNTIAN","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2607.01239&json=true","fetch_graph":"https://pith.science/api/pith-number/GDJNTIANT4W5J2X7ZCFGDWNS3W/graph.json","fetch_events":"https://pith.science/api/pith-number/GDJNTIANT4W5J2X7ZCFGDWNS3W/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/GDJNTIANT4W5J2X7ZCFGDWNS3W/action/timestamp_anchor","attest_storage":"https://pith.science/pith/GDJNTIANT4W5J2X7ZCFGDWNS3W/action/storage_attestation","attest_author":"https://pith.science/pith/GDJNTIANT4W5J2X7ZCFGDWNS3W/action/author_attestation","sign_citation":"https://pith.science/pith/GDJNTIANT4W5J2X7ZCFGDWNS3W/action/citation_signature","submit_replication":"https://pith.science/pith/GDJNTIANT4W5J2X7ZCFGDWNS3W/action/replication_record"}},"created_at":"2026-07-03T00:16:55.053436+00:00","updated_at":"2026-07-03T00:16:55.053436+00:00"}