{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:6HDFXMHIPD3XAR7EEC5O47BPLL","short_pith_number":"pith:6HDFXMHI","schema_version":"1.0","canonical_sha256":"f1c65bb0e878f77047e420baee7c2f5ac2e377f36c44c985b93275bba0000938","source":{"kind":"arxiv","id":"2605.27851","version":1},"attestation_state":"computed","paper":{"title":"When Context Flips, Safety Breaks: Diagnosing Brittle Safety in Aligned Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Alex Kwon, Dasol Choi","submitted_at":"2026-05-27T02:09:30Z","abstract_excerpt":"Safety benchmark scores provide incomplete evidence of deployment readiness: aligned language models often adhere to rigid rules even when a situational update flips which action is safe. We term this failure brittle safety. To diagnose it, we introduce context-flip evaluation, testing 12 models across a safety benchmark (PacifAIst) and two commonsense controls using paired variants where the nominally safe action produces harm. Three findings emerge. First, brittle safety is safety-specific: all 12 models exhibit a safety-commonsense gap (mean +17.4 pp). Baseline accuracy fails to predict bri"},"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":"2605.27851","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2026-05-27T02:09:30Z","cross_cats_sorted":[],"title_canon_sha256":"d7a59eb248b87df42335842b72a66677cd3c684ded7602bc2a5a9f7e0229e72a","abstract_canon_sha256":"3bfa80399d49fe7c08c1bd7484e2218aa47eb09bdb5737235f377f7ed86242fe"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-28T01:04:50.442119Z","signature_b64":"doHvklv3+rBBQDVXdvjC5zNzd/dPhhx1Q35VqWSVlHtfYgFbgEi4kOyJFbpqyClP7VPBOQ8lPgoneaE68+S2DA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"f1c65bb0e878f77047e420baee7c2f5ac2e377f36c44c985b93275bba0000938","last_reissued_at":"2026-05-28T01:04:50.441731Z","signature_status":"signed_v1","first_computed_at":"2026-05-28T01:04:50.441731Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"When Context Flips, Safety Breaks: Diagnosing Brittle Safety in Aligned Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Alex Kwon, Dasol Choi","submitted_at":"2026-05-27T02:09:30Z","abstract_excerpt":"Safety benchmark scores provide incomplete evidence of deployment readiness: aligned language models often adhere to rigid rules even when a situational update flips which action is safe. We term this failure brittle safety. To diagnose it, we introduce context-flip evaluation, testing 12 models across a safety benchmark (PacifAIst) and two commonsense controls using paired variants where the nominally safe action produces harm. Three findings emerge. First, brittle safety is safety-specific: all 12 models exhibit a safety-commonsense gap (mean +17.4 pp). Baseline accuracy fails to predict bri"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.27851","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/2605.27851/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":"2605.27851","created_at":"2026-05-28T01:04:50.441786+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.27851v1","created_at":"2026-05-28T01:04:50.441786+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.27851","created_at":"2026-05-28T01:04:50.441786+00:00"},{"alias_kind":"pith_short_12","alias_value":"6HDFXMHIPD3X","created_at":"2026-05-28T01:04:50.441786+00:00"},{"alias_kind":"pith_short_16","alias_value":"6HDFXMHIPD3XAR7E","created_at":"2026-05-28T01:04:50.441786+00:00"},{"alias_kind":"pith_short_8","alias_value":"6HDFXMHI","created_at":"2026-05-28T01:04:50.441786+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2606.29279","citing_title":"Manufactured Confidence: How Memory Consolidation Turns Hearsay into Confident Facts","ref_index":22,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/6HDFXMHIPD3XAR7EEC5O47BPLL","json":"https://pith.science/pith/6HDFXMHIPD3XAR7EEC5O47BPLL.json","graph_json":"https://pith.science/api/pith-number/6HDFXMHIPD3XAR7EEC5O47BPLL/graph.json","events_json":"https://pith.science/api/pith-number/6HDFXMHIPD3XAR7EEC5O47BPLL/events.json","paper":"https://pith.science/paper/6HDFXMHI"},"agent_actions":{"view_html":"https://pith.science/pith/6HDFXMHIPD3XAR7EEC5O47BPLL","download_json":"https://pith.science/pith/6HDFXMHIPD3XAR7EEC5O47BPLL.json","view_paper":"https://pith.science/paper/6HDFXMHI","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.27851&json=true","fetch_graph":"https://pith.science/api/pith-number/6HDFXMHIPD3XAR7EEC5O47BPLL/graph.json","fetch_events":"https://pith.science/api/pith-number/6HDFXMHIPD3XAR7EEC5O47BPLL/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/6HDFXMHIPD3XAR7EEC5O47BPLL/action/timestamp_anchor","attest_storage":"https://pith.science/pith/6HDFXMHIPD3XAR7EEC5O47BPLL/action/storage_attestation","attest_author":"https://pith.science/pith/6HDFXMHIPD3XAR7EEC5O47BPLL/action/author_attestation","sign_citation":"https://pith.science/pith/6HDFXMHIPD3XAR7EEC5O47BPLL/action/citation_signature","submit_replication":"https://pith.science/pith/6HDFXMHIPD3XAR7EEC5O47BPLL/action/replication_record"}},"created_at":"2026-05-28T01:04:50.441786+00:00","updated_at":"2026-05-28T01:04:50.441786+00:00"}