{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:UBFUKC2Y7357I3TFO4T5E2BZIS","short_pith_number":"pith:UBFUKC2Y","schema_version":"1.0","canonical_sha256":"a04b450b58fefbf46e657727d2683944b1d63e9759971b6a8ef3fe82eb0866d4","source":{"kind":"arxiv","id":"2502.14744","version":4},"attestation_state":"computed","paper":{"title":"HiddenDetect: Detecting Jailbreak Attacks against Large Vision-Language Models via Monitoring Hidden States","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Bo Zheng, Tianshuo Peng, Xiangyu Yue, Xiaoyong Zhu, Xinyan Gao, Yilei Jiang, Yingshui Tan","submitted_at":"2025-02-20T17:14:34Z","abstract_excerpt":"The integration of additional modalities increases the susceptibility of large vision-language models (LVLMs) to safety risks, such as jailbreak attacks, compared to their language-only counterparts. While existing research primarily focuses on post-hoc alignment techniques, the underlying safety mechanisms within LVLMs remain largely unexplored. In this work , we investigate whether LVLMs inherently encode safety-relevant signals within their internal activations during inference. Our findings reveal that LVLMs exhibit distinct activation patterns when processing unsafe prompts, which can be "},"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":"2502.14744","kind":"arxiv","version":4},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-02-20T17:14:34Z","cross_cats_sorted":[],"title_canon_sha256":"a9b4c048e7df139f9ae2eb381ca54f59e8070476da50ab3227ed4c8aca92b2ed","abstract_canon_sha256":"db5eb6f04253c05be7a1b322a5590c2a8744b1cae8e1bc5cf1f0ebdf0e50bfae"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:25:42.249074Z","signature_b64":"ym3NcRj0h5ps3gJySIaSX3EUSe8hMJzAO9fFK8eYH1ehAQv8ytwU4cqX3zWBIYp/iQnAeu6FSlt+7WcfNHB0Cw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a04b450b58fefbf46e657727d2683944b1d63e9759971b6a8ef3fe82eb0866d4","last_reissued_at":"2026-07-05T11:25:42.248591Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:25:42.248591Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"HiddenDetect: Detecting Jailbreak Attacks against Large Vision-Language Models via Monitoring Hidden States","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Bo Zheng, Tianshuo Peng, Xiangyu Yue, Xiaoyong Zhu, Xinyan Gao, Yilei Jiang, Yingshui Tan","submitted_at":"2025-02-20T17:14:34Z","abstract_excerpt":"The integration of additional modalities increases the susceptibility of large vision-language models (LVLMs) to safety risks, such as jailbreak attacks, compared to their language-only counterparts. While existing research primarily focuses on post-hoc alignment techniques, the underlying safety mechanisms within LVLMs remain largely unexplored. In this work , we investigate whether LVLMs inherently encode safety-relevant signals within their internal activations during inference. Our findings reveal that LVLMs exhibit distinct activation patterns when processing unsafe prompts, which can be "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.14744","kind":"arxiv","version":4},"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/2502.14744/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":"2502.14744","created_at":"2026-07-05T11:25:42.248655+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.14744v4","created_at":"2026-07-05T11:25:42.248655+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.14744","created_at":"2026-07-05T11:25:42.248655+00:00"},{"alias_kind":"pith_short_12","alias_value":"UBFUKC2Y7357","created_at":"2026-07-05T11:25:42.248655+00:00"},{"alias_kind":"pith_short_16","alias_value":"UBFUKC2Y7357I3TF","created_at":"2026-07-05T11:25:42.248655+00:00"},{"alias_kind":"pith_short_8","alias_value":"UBFUKC2Y","created_at":"2026-07-05T11:25:42.248655+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.07335","citing_title":"Defending Jailbreak Attacks on Large Language Models via Manifold Trajectory Kinetics","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09391","citing_title":"Do Linear Probes Generalize Better in Persona Coordinates?","ref_index":24,"is_internal_anchor":false},{"citing_arxiv_id":"2511.17408","citing_title":"The Impact of Off-Policy Training Data on Probe Generalisation","ref_index":21,"is_internal_anchor":false},{"citing_arxiv_id":"2605.03095","citing_title":"Revisiting JBShield: Breaking and Rebuilding Representation-Level Jailbreak Defenses","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09391","citing_title":"Do Linear Probes Generalize Better in Persona Coordinates?","ref_index":24,"is_internal_anchor":false},{"citing_arxiv_id":"2604.06247","citing_title":"SALLIE: Safeguarding Against Latent Language & Image Exploits","ref_index":11,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/UBFUKC2Y7357I3TFO4T5E2BZIS","json":"https://pith.science/pith/UBFUKC2Y7357I3TFO4T5E2BZIS.json","graph_json":"https://pith.science/api/pith-number/UBFUKC2Y7357I3TFO4T5E2BZIS/graph.json","events_json":"https://pith.science/api/pith-number/UBFUKC2Y7357I3TFO4T5E2BZIS/events.json","paper":"https://pith.science/paper/UBFUKC2Y"},"agent_actions":{"view_html":"https://pith.science/pith/UBFUKC2Y7357I3TFO4T5E2BZIS","download_json":"https://pith.science/pith/UBFUKC2Y7357I3TFO4T5E2BZIS.json","view_paper":"https://pith.science/paper/UBFUKC2Y","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.14744&json=true","fetch_graph":"https://pith.science/api/pith-number/UBFUKC2Y7357I3TFO4T5E2BZIS/graph.json","fetch_events":"https://pith.science/api/pith-number/UBFUKC2Y7357I3TFO4T5E2BZIS/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/UBFUKC2Y7357I3TFO4T5E2BZIS/action/timestamp_anchor","attest_storage":"https://pith.science/pith/UBFUKC2Y7357I3TFO4T5E2BZIS/action/storage_attestation","attest_author":"https://pith.science/pith/UBFUKC2Y7357I3TFO4T5E2BZIS/action/author_attestation","sign_citation":"https://pith.science/pith/UBFUKC2Y7357I3TFO4T5E2BZIS/action/citation_signature","submit_replication":"https://pith.science/pith/UBFUKC2Y7357I3TFO4T5E2BZIS/action/replication_record"}},"created_at":"2026-07-05T11:25:42.248655+00:00","updated_at":"2026-07-05T11:25:42.248655+00:00"}