{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:4SZQPFNMFYPDT5UY7KJGXSTMPA","short_pith_number":"pith:4SZQPFNM","schema_version":"1.0","canonical_sha256":"e4b30795ac2e1e39f698fa926bca6c780402fc7894fc57fd6073e190c2da87af","source":{"kind":"arxiv","id":"2607.14147","version":1},"attestation_state":"computed","paper":{"title":"Breaking Refusal in the First Half: A Mechanistic Study of the Prefill Jailbreak","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CR","cs.LG"],"primary_cat":"cs.CL","authors_text":"Alex Kwon","submitted_at":"2026-07-14T09:30:34Z","abstract_excerpt":"Aligned language models refuse harmful requests, but a one-line prefill (\"Sure, here is\") strips the refusal. We ask where and how it fails. The harm representation stays intact: on the prompts the attack flips to compliance, a linear probe reads harm as high as on the refused ones (0.91-0.98), while behavioral refusal drops to chance. This holds across four models and three families (1.5-3.8B, and at 14B). Refusal is therefore a shallow, response-site computation. We localize it to an early window: a dose-matched position control shows the first half of the response suffices to break refusal,"},"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":"2607.14147","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-07-14T09:30:34Z","cross_cats_sorted":["cs.AI","cs.CR","cs.LG"],"title_canon_sha256":"9bdf0b1ed0419e54a1fab33f2f5bcf79dd653047d03984737a30797ad82e7068","abstract_canon_sha256":"42b3d5d52b7ab6d49570f45b696acfa5e1f8b5971ca9b383873d98e7cd60520e"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-17T00:20:54.003178Z","signature_b64":"X51EoXUAgqNEq0/v/j+jI9agExTSzux4qyiup7BkQYqffSRK/1qCVwJG1d/4HJshB4L0jHBe8xZDzfjeUp40Bg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"e4b30795ac2e1e39f698fa926bca6c780402fc7894fc57fd6073e190c2da87af","last_reissued_at":"2026-07-17T00:20:54.002328Z","signature_status":"signed_v1","first_computed_at":"2026-07-17T00:20:54.002328Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Breaking Refusal in the First Half: A Mechanistic Study of the Prefill Jailbreak","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CR","cs.LG"],"primary_cat":"cs.CL","authors_text":"Alex Kwon","submitted_at":"2026-07-14T09:30:34Z","abstract_excerpt":"Aligned language models refuse harmful requests, but a one-line prefill (\"Sure, here is\") strips the refusal. We ask where and how it fails. The harm representation stays intact: on the prompts the attack flips to compliance, a linear probe reads harm as high as on the refused ones (0.91-0.98), while behavioral refusal drops to chance. This holds across four models and three families (1.5-3.8B, and at 14B). Refusal is therefore a shallow, response-site computation. We localize it to an early window: a dose-matched position control shows the first half of the response suffices to break refusal,"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2607.14147","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/2607.14147/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":"2607.14147","created_at":"2026-07-17T00:20:54.002723+00:00"},{"alias_kind":"arxiv_version","alias_value":"2607.14147v1","created_at":"2026-07-17T00:20:54.002723+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2607.14147","created_at":"2026-07-17T00:20:54.002723+00:00"},{"alias_kind":"pith_short_12","alias_value":"4SZQPFNMFYPD","created_at":"2026-07-17T00:20:54.002723+00:00"},{"alias_kind":"pith_short_16","alias_value":"4SZQPFNMFYPDT5UY","created_at":"2026-07-17T00:20:54.002723+00:00"},{"alias_kind":"pith_short_8","alias_value":"4SZQPFNM","created_at":"2026-07-17T00:20:54.002723+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/4SZQPFNMFYPDT5UY7KJGXSTMPA","json":"https://pith.science/pith/4SZQPFNMFYPDT5UY7KJGXSTMPA.json","graph_json":"https://pith.science/api/pith-number/4SZQPFNMFYPDT5UY7KJGXSTMPA/graph.json","events_json":"https://pith.science/api/pith-number/4SZQPFNMFYPDT5UY7KJGXSTMPA/events.json","paper":"https://pith.science/paper/4SZQPFNM"},"agent_actions":{"view_html":"https://pith.science/pith/4SZQPFNMFYPDT5UY7KJGXSTMPA","download_json":"https://pith.science/pith/4SZQPFNMFYPDT5UY7KJGXSTMPA.json","view_paper":"https://pith.science/paper/4SZQPFNM","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2607.14147&json=true","fetch_graph":"https://pith.science/api/pith-number/4SZQPFNMFYPDT5UY7KJGXSTMPA/graph.json","fetch_events":"https://pith.science/api/pith-number/4SZQPFNMFYPDT5UY7KJGXSTMPA/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/4SZQPFNMFYPDT5UY7KJGXSTMPA/action/timestamp_anchor","attest_storage":"https://pith.science/pith/4SZQPFNMFYPDT5UY7KJGXSTMPA/action/storage_attestation","attest_author":"https://pith.science/pith/4SZQPFNMFYPDT5UY7KJGXSTMPA/action/author_attestation","sign_citation":"https://pith.science/pith/4SZQPFNMFYPDT5UY7KJGXSTMPA/action/citation_signature","submit_replication":"https://pith.science/pith/4SZQPFNMFYPDT5UY7KJGXSTMPA/action/replication_record"}},"created_at":"2026-07-17T00:20:54.002723+00:00","updated_at":"2026-07-17T00:20:54.002723+00:00"}