{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2026:KIOBQZTCK252DPBZ4XBSY7P6FQ","short_pith_number":"pith:KIOBQZTC","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"},"canonical_sha256":"521c18666256bba1bc39e5c32c7dfe2c20c2c3843ec55fa1317115eb92f9f7d7","source":{"kind":"arxiv","id":"2605.12726","version":1},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.12726","created_at":"2026-05-18T03:09:49Z"},{"alias_kind":"arxiv_version","alias_value":"2605.12726v1","created_at":"2026-05-18T03:09:49Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.12726","created_at":"2026-05-18T03:09:49Z"},{"alias_kind":"pith_short_12","alias_value":"KIOBQZTCK252","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"KIOBQZTCK252DPBZ","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"KIOBQZTC","created_at":"2026-05-18T12:33:37Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2026:KIOBQZTCK252DPBZ4XBSY7P6FQ","target":"record","payload":{"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"},"canonical_sha256":"521c18666256bba1bc39e5c32c7dfe2c20c2c3843ec55fa1317115eb92f9f7d7","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"},"source_kind":"arxiv","source_id":"2605.12726","source_version":1,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-18T03:09:49Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"Dwb9w1Nb+HktbPUiAsMa388V+M9onTh2bVlll3l7btcj0WkLgVOEMC+DbeGsRDv4KY0wQyYUZ3P86khYiS3wDA==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-08-04T14:47:59.337844Z"},"content_sha256":"37908236817e3aea6a94988f165090be6869052baf40034c7e1fffa98c49caef","schema_version":"1.0","event_id":"sha256:37908236817e3aea6a94988f165090be6869052baf40034c7e1fffa98c49caef"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2026:KIOBQZTCK252DPBZ4XBSY7P6FQ","target":"graph","payload":{"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"},"verdict_id":"661049f7-796b-4572-b51e-4f9023b33689"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-18T03:09:49Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"Va0gdegkxWVRvhTensOL18257WcOxGdVCidbspZ+e5wH5DY/grAhg2KkcRnywkiPHbhEcAs/Pnzx+OPPmin1Bg==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-08-04T14:47:59.338683Z"},"content_sha256":"2ee807c06f2dd8656ff8130b019247a012eccc2288e1f4b4f808579b3cba203e","schema_version":"1.0","event_id":"sha256:2ee807c06f2dd8656ff8130b019247a012eccc2288e1f4b4f808579b3cba203e"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/KIOBQZTCK252DPBZ4XBSY7P6FQ/bundle.json","state_url":"https://pith.science/pith/KIOBQZTCK252DPBZ4XBSY7P6FQ/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/KIOBQZTCK252DPBZ4XBSY7P6FQ/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-08-04T14:47:59Z","links":{"resolver":"https://pith.science/pith/KIOBQZTCK252DPBZ4XBSY7P6FQ","bundle":"https://pith.science/pith/KIOBQZTCK252DPBZ4XBSY7P6FQ/bundle.json","state":"https://pith.science/pith/KIOBQZTCK252DPBZ4XBSY7P6FQ/state.json","well_known_bundle":"https://pith.science/.well-known/pith/KIOBQZTCK252DPBZ4XBSY7P6FQ/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:KIOBQZTCK252DPBZ4XBSY7P6FQ","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"2cc9767e40cfc68be3d27a2a0fecf791b522829c07d9270ddcf669032f24201c","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-12T20:30:24Z","title_canon_sha256":"7f878d548d0a924a9b6b0a69ccd24f295e112830e1b2eaa51cc49b99a58870f8"},"schema_version":"1.0","source":{"id":"2605.12726","kind":"arxiv","version":1}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.12726","created_at":"2026-05-18T03:09:49Z"},{"alias_kind":"arxiv_version","alias_value":"2605.12726v1","created_at":"2026-05-18T03:09:49Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.12726","created_at":"2026-05-18T03:09:49Z"},{"alias_kind":"pith_short_12","alias_value":"KIOBQZTCK252","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"KIOBQZTCK252DPBZ","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"KIOBQZTC","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:2ee807c06f2dd8656ff8130b019247a012eccc2288e1f4b4f808579b3cba203e","target":"graph","created_at":"2026-05-18T03:09:49Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","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."}],"snapshot_sha256":"509465bd0b86264ac9129bce6e6168ee58dd0ac572ced44ac3572849e7440c84"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"paper":{"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 ","authors_text":"Shravan Doda","cross_cats":[],"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.","license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-12T20:30:24Z","title":"Before the Last Token: Diagnosing Final-Token Safety Probe Failures"},"references":{"count":16,"internal_anchors":7,"resolved_work":16,"sample":[{"cited_arxiv_id":"2406.11717","doi":"","is_internal_anchor":true,"ref_index":1,"title":"Refusal in Language Models Is Mediated by a Single Direction","work_id":"fbb9538d-8e58-4902-9fbd-b11f044bc2d5","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"Safeswitch: Steering unsafe llm behavior via internal activation signals","work_id":"d01742b8-94af-4998-8813-ba590a1e0b3c","year":null},{"cited_arxiv_id":"2310.06825","doi":"","is_internal_anchor":true,"ref_index":3,"title":"Lin, Z., Yang, J., Qiu, Y ., Guo, H., Bao, Y ., and Guan, Y","work_id":"eb5e1305-ad11-4875-ad8d-ad8b8f697599","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":4,"title":"org/abs/2511.14195","work_id":"bbd4bb60-60bb-424e-8e67-24e1f5742483","year":null},{"cited_arxiv_id":"2604.07727","doi":"","is_internal_anchor":true,"ref_index":5,"title":"TrajGuard: Streaming Hidden-state Trajectory Detection for Decoding-time Jailbreak Defense","work_id":"81cce937-01b1-45b6-adfc-d4a0e5a1be24","year":null}],"snapshot_sha256":"da5ba29450b1da15aeb46357c3f20c7cd5126e042182532288d73188af8da3db"},"source":{"id":"2605.12726","kind":"arxiv","version":1},"verdict":{"created_at":"2026-05-14T21:33:49.301876Z","id":"661049f7-796b-4572-b51e-4f9023b33689","model_set":{"reader":"grok-4.3"},"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","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.","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.","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."}},"verdict_id":"661049f7-796b-4572-b51e-4f9023b33689"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:37908236817e3aea6a94988f165090be6869052baf40034c7e1fffa98c49caef","target":"record","created_at":"2026-05-18T03:09:49Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"2cc9767e40cfc68be3d27a2a0fecf791b522829c07d9270ddcf669032f24201c","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-12T20:30:24Z","title_canon_sha256":"7f878d548d0a924a9b6b0a69ccd24f295e112830e1b2eaa51cc49b99a58870f8"},"schema_version":"1.0","source":{"id":"2605.12726","kind":"arxiv","version":1}},"canonical_sha256":"521c18666256bba1bc39e5c32c7dfe2c20c2c3843ec55fa1317115eb92f9f7d7","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"521c18666256bba1bc39e5c32c7dfe2c20c2c3843ec55fa1317115eb92f9f7d7","first_computed_at":"2026-05-18T03:09:49.321911Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-18T03:09:49.321911Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"inXlRK/Mo7xifHbCH9g4sxl6K08rBePp9g6WKTSMxppNVNY4IKLRmqUrup2oPY7nUxUObqSBrd5/kQiSTR/dCw==","signature_status":"signed_v1","signed_at":"2026-05-18T03:09:49.322744Z","signed_message":"canonical_sha256_bytes"},"source_id":"2605.12726","source_kind":"arxiv","source_version":1}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:37908236817e3aea6a94988f165090be6869052baf40034c7e1fffa98c49caef","sha256:2ee807c06f2dd8656ff8130b019247a012eccc2288e1f4b4f808579b3cba203e"],"state_sha256":"e1853e7a73dccd25ff5a41fe06cd45717ab8661037c18503bef6b1c581c57a48"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"SO56NiJaoHoNuo8Ia+OlH00v4v8O/iTWZBO0vYQn8ZiFnzRhIctZpQYOZl2mUeR1P7dNwmbkuK5jcLcJrb/ZBg==","signed_message":"bundle_sha256_bytes","signed_at":"2026-08-04T14:47:59.344194Z","bundle_sha256":"f97df74a4d50489d0bba05468e76b51d9a6f137cc2065f6015104911ea912d01"}}