{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:KIOBQZTCK252DPBZ4XBSY7P6FQ","short_pith_number":"pith:KIOBQZTC","schema_version":"1.0","canonical_sha256":"521c18666256bba1bc39e5c32c7dfe2c20c2c3843ec55fa1317115eb92f9f7d7","source":{"kind":"arxiv","id":"2605.12726","version":1},"attestation_state":"computed","paper":{"title":"Before the Last Token: Diagnosing Final-Token Safety Probe Failures","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Final-token safety probes miss jailbreak evidence spread across earlier tokens, but a PCA-HMM model on prefill trajectories recovers many such cases without high false positives.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Shravan Doda","submitted_at":"2026-05-12T20:30:24Z","abstract_excerpt":"Final-token safety probes monitor a single hidden state after prompt prefill, but jailbreak prompts can contain probe-visible unsafe evidence distributed across earlier user-token representations that is missed by this readout. We study this prefill-time failure mode using SafeSwitch-style probes trained only on clean harmful and benign prompts across three instruction-tuned LLMs. The probes achieve high recall on clean harmful prompts, but miss many jailbreaks and can produce false positives on safety-adjacent benign prompts. Subspace analyses suggest that missed jailbreaks differ from clean "},"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":true,"formal_links_present":false},"canonical_record":{"source":{"id":"2605.12726","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-12T20:30:24Z","cross_cats_sorted":[],"title_canon_sha256":"7f878d548d0a924a9b6b0a69ccd24f295e112830e1b2eaa51cc49b99a58870f8","abstract_canon_sha256":"2cc9767e40cfc68be3d27a2a0fecf791b522829c07d9270ddcf669032f24201c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-18T03:09:49.322744Z","signature_b64":"inXlRK/Mo7xifHbCH9g4sxl6K08rBePp9g6WKTSMxppNVNY4IKLRmqUrup2oPY7nUxUObqSBrd5/kQiSTR/dCw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"521c18666256bba1bc39e5c32c7dfe2c20c2c3843ec55fa1317115eb92f9f7d7","last_reissued_at":"2026-05-18T03:09:49.321911Z","signature_status":"signed_v1","first_computed_at":"2026-05-18T03:09:49.321911Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Before the Last Token: Diagnosing Final-Token Safety Probe Failures","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Final-token safety probes miss jailbreak evidence spread across earlier tokens, but a PCA-HMM model on prefill trajectories recovers many such cases without high false positives.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Shravan Doda","submitted_at":"2026-05-12T20:30:24Z","abstract_excerpt":"Final-token safety probes monitor a single hidden state after prompt prefill, but jailbreak prompts can contain probe-visible unsafe evidence distributed across earlier user-token representations that is missed by this readout. We study this prefill-time failure mode using SafeSwitch-style probes trained only on clean harmful and benign prompts across three instruction-tuned LLMs. The probes achieve high recall on clean harmful prompts, but miss many jailbreaks and can produce false positives on safety-adjacent benign prompts. Subspace analyses suggest that missed jailbreaks differ from clean "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"A simple PCA-HMM trajectory model, trained only on the same clean split, recovers many final-token misses from user-content prefill trajectories without the catastrophic false-positive behavior of naive token pooling.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That subspace analyses accurately identify the directions missed by the probe and that the PCA-HMM model trained on clean prompts generalizes to recover jailbreak cases without introducing new failure modes.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Final-token probes miss distributed unsafe evidence in jailbreaks, but a PCA-HMM model on prefill trajectories recovers many misses without naive pooling's false positives.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Final-token safety probes miss jailbreak evidence spread across earlier tokens, but a PCA-HMM model on prefill trajectories recovers many such cases without high false positives.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"509465bd0b86264ac9129bce6e6168ee58dd0ac572ced44ac3572849e7440c84"},"source":{"id":"2605.12726","kind":"arxiv","version":1},"verdict":{"id":"661049f7-796b-4572-b51e-4f9023b33689","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T21:33:49.301876Z","strongest_claim":"A simple PCA-HMM trajectory model, trained only on the same clean split, recovers many final-token misses from user-content prefill trajectories without the catastrophic false-positive behavior of naive token pooling.","one_line_summary":"Final-token probes miss distributed unsafe evidence in jailbreaks, but a PCA-HMM model on prefill trajectories recovers many misses without naive pooling's false positives.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That subspace analyses accurately identify the directions missed by the probe and that the PCA-HMM model trained on clean prompts generalizes to recover jailbreak cases without introducing new failure modes.","pith_extraction_headline":"Final-token safety probes miss jailbreak evidence spread across earlier tokens, but a PCA-HMM model on prefill trajectories recovers many such cases without high false positives."},"references":{"count":16,"sample":[{"doi":"","year":null,"title":"Refusal in Language Models Is Mediated by a Single Direction","work_id":"fbb9538d-8e58-4902-9fbd-b11f044bc2d5","ref_index":1,"cited_arxiv_id":"2406.11717","is_internal_anchor":true},{"doi":"","year":null,"title":"Safeswitch: Steering unsafe llm behavior via internal activation signals","work_id":"d01742b8-94af-4998-8813-ba590a1e0b3c","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Lin, Z., Yang, J., Qiu, Y ., Guo, H., Bao, Y ., and Guan, Y","work_id":"eb5e1305-ad11-4875-ad8d-ad8b8f697599","ref_index":3,"cited_arxiv_id":"2310.06825","is_internal_anchor":true},{"doi":"","year":null,"title":"org/abs/2511.14195","work_id":"bbd4bb60-60bb-424e-8e67-24e1f5742483","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"TrajGuard: Streaming Hidden-state Trajectory Detection for Decoding-time Jailbreak Defense","work_id":"81cce937-01b1-45b6-adfc-d4a0e5a1be24","ref_index":5,"cited_arxiv_id":"2604.07727","is_internal_anchor":true}],"resolved_work":16,"snapshot_sha256":"da5ba29450b1da15aeb46357c3f20c7cd5126e042182532288d73188af8da3db","internal_anchors":7},"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":"2605.12726","created_at":"2026-05-18T03:09:49.322036+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.12726v1","created_at":"2026-05-18T03:09:49.322036+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.12726","created_at":"2026-05-18T03:09:49.322036+00:00"},{"alias_kind":"pith_short_12","alias_value":"KIOBQZTCK252","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"KIOBQZTCK252DPBZ","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"KIOBQZTC","created_at":"2026-05-18T12:33:37.589309+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/KIOBQZTCK252DPBZ4XBSY7P6FQ","json":"https://pith.science/pith/KIOBQZTCK252DPBZ4XBSY7P6FQ.json","graph_json":"https://pith.science/api/pith-number/KIOBQZTCK252DPBZ4XBSY7P6FQ/graph.json","events_json":"https://pith.science/api/pith-number/KIOBQZTCK252DPBZ4XBSY7P6FQ/events.json","paper":"https://pith.science/paper/KIOBQZTC"},"agent_actions":{"view_html":"https://pith.science/pith/KIOBQZTCK252DPBZ4XBSY7P6FQ","download_json":"https://pith.science/pith/KIOBQZTCK252DPBZ4XBSY7P6FQ.json","view_paper":"https://pith.science/paper/KIOBQZTC","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.12726&json=true","fetch_graph":"https://pith.science/api/pith-number/KIOBQZTCK252DPBZ4XBSY7P6FQ/graph.json","fetch_events":"https://pith.science/api/pith-number/KIOBQZTCK252DPBZ4XBSY7P6FQ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/KIOBQZTCK252DPBZ4XBSY7P6FQ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/KIOBQZTCK252DPBZ4XBSY7P6FQ/action/storage_attestation","attest_author":"https://pith.science/pith/KIOBQZTCK252DPBZ4XBSY7P6FQ/action/author_attestation","sign_citation":"https://pith.science/pith/KIOBQZTCK252DPBZ4XBSY7P6FQ/action/citation_signature","submit_replication":"https://pith.science/pith/KIOBQZTCK252DPBZ4XBSY7P6FQ/action/replication_record"}},"created_at":"2026-05-18T03:09:49.322036+00:00","updated_at":"2026-05-18T03:09:49.322036+00:00"}