{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:6U5A3QHG3BWA6DJUOYFMYYAHCE","short_pith_number":"pith:6U5A3QHG","schema_version":"1.0","canonical_sha256":"f53a0dc0e6d86c0f0d34760acc6007111678d9a8ab10c07d3bb62638d1fa4611","source":{"kind":"arxiv","id":"2507.04365","version":1},"attestation_state":"computed","paper":{"title":"Attention Slipping: A Mechanistic Understanding of Jailbreak Attacks and Defenses in LLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.CR","authors_text":"Pin-Yu Chen, Tsung-Yi Ho, Xiaomeng Hu","submitted_at":"2025-07-06T12:19:04Z","abstract_excerpt":"As large language models (LLMs) become more integral to society and technology, ensuring their safety becomes essential. Jailbreak attacks exploit vulnerabilities to bypass safety guardrails, posing a significant threat. However, the mechanisms enabling these attacks are not well understood. In this paper, we reveal a universal phenomenon that occurs during jailbreak attacks: Attention Slipping. During this phenomenon, the model gradually reduces the attention it allocates to unsafe requests in a user query during the attack process, ultimately causing a jailbreak. We show Attention Slipping i"},"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":"2507.04365","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CR","submitted_at":"2025-07-06T12:19:04Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"ab02d8bb814720a7b226e305c14fb0ab888920f9227e7c9a64c2c524ba9505d2","abstract_canon_sha256":"0fdaf3152a837e042551840a625a08fdb24c0384c402ad11b19a323e0229be31"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:32:46.548157Z","signature_b64":"9zA5noWr1mztLk9sy4O5eymIgL3evzaYWt/lrf13oI1aV0LrW7ObNOWSGDv/yH0Z1aBWFblI+qgnqYbo2QEqAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"f53a0dc0e6d86c0f0d34760acc6007111678d9a8ab10c07d3bb62638d1fa4611","last_reissued_at":"2026-07-05T11:32:46.547783Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:32:46.547783Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Attention Slipping: A Mechanistic Understanding of Jailbreak Attacks and Defenses in LLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.CR","authors_text":"Pin-Yu Chen, Tsung-Yi Ho, Xiaomeng Hu","submitted_at":"2025-07-06T12:19:04Z","abstract_excerpt":"As large language models (LLMs) become more integral to society and technology, ensuring their safety becomes essential. Jailbreak attacks exploit vulnerabilities to bypass safety guardrails, posing a significant threat. However, the mechanisms enabling these attacks are not well understood. In this paper, we reveal a universal phenomenon that occurs during jailbreak attacks: Attention Slipping. During this phenomenon, the model gradually reduces the attention it allocates to unsafe requests in a user query during the attack process, ultimately causing a jailbreak. We show Attention Slipping i"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2507.04365","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/2507.04365/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":"2507.04365","created_at":"2026-07-05T11:32:46.547839+00:00"},{"alias_kind":"arxiv_version","alias_value":"2507.04365v1","created_at":"2026-07-05T11:32:46.547839+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2507.04365","created_at":"2026-07-05T11:32:46.547839+00:00"},{"alias_kind":"pith_short_12","alias_value":"6U5A3QHG3BWA","created_at":"2026-07-05T11:32:46.547839+00:00"},{"alias_kind":"pith_short_16","alias_value":"6U5A3QHG3BWA6DJU","created_at":"2026-07-05T11:32:46.547839+00:00"},{"alias_kind":"pith_short_8","alias_value":"6U5A3QHG","created_at":"2026-07-05T11:32:46.547839+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.05609","citing_title":"SlotGCG: Exploiting the Positional Vulnerability in LLMs for Jailbreak Attacks","ref_index":21,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/6U5A3QHG3BWA6DJUOYFMYYAHCE","json":"https://pith.science/pith/6U5A3QHG3BWA6DJUOYFMYYAHCE.json","graph_json":"https://pith.science/api/pith-number/6U5A3QHG3BWA6DJUOYFMYYAHCE/graph.json","events_json":"https://pith.science/api/pith-number/6U5A3QHG3BWA6DJUOYFMYYAHCE/events.json","paper":"https://pith.science/paper/6U5A3QHG"},"agent_actions":{"view_html":"https://pith.science/pith/6U5A3QHG3BWA6DJUOYFMYYAHCE","download_json":"https://pith.science/pith/6U5A3QHG3BWA6DJUOYFMYYAHCE.json","view_paper":"https://pith.science/paper/6U5A3QHG","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2507.04365&json=true","fetch_graph":"https://pith.science/api/pith-number/6U5A3QHG3BWA6DJUOYFMYYAHCE/graph.json","fetch_events":"https://pith.science/api/pith-number/6U5A3QHG3BWA6DJUOYFMYYAHCE/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/6U5A3QHG3BWA6DJUOYFMYYAHCE/action/timestamp_anchor","attest_storage":"https://pith.science/pith/6U5A3QHG3BWA6DJUOYFMYYAHCE/action/storage_attestation","attest_author":"https://pith.science/pith/6U5A3QHG3BWA6DJUOYFMYYAHCE/action/author_attestation","sign_citation":"https://pith.science/pith/6U5A3QHG3BWA6DJUOYFMYYAHCE/action/citation_signature","submit_replication":"https://pith.science/pith/6U5A3QHG3BWA6DJUOYFMYYAHCE/action/replication_record"}},"created_at":"2026-07-05T11:32:46.547839+00:00","updated_at":"2026-07-05T11:32:46.547839+00:00"}