{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:C44PCWVMADHSUO3LNN5V52GKD6","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":"1e5858e07c56143bc8ec3894c094d9d5073d73193f6f9129aec9b301cb525788","cross_cats_sorted":["cs.AI"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-01-07T20:53:18Z","title_canon_sha256":"cc28d64672e15f525a89530e71aa2adb9c50a1d166cde7c89ec0ff0bf76754b6"},"schema_version":"1.0","source":{"id":"2601.04389","kind":"arxiv","version":3}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2601.04389","created_at":"2026-06-23T02:13:19Z"},{"alias_kind":"arxiv_version","alias_value":"2601.04389v3","created_at":"2026-06-23T02:13:19Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2601.04389","created_at":"2026-06-23T02:13:19Z"},{"alias_kind":"pith_short_12","alias_value":"C44PCWVMADHS","created_at":"2026-06-23T02:13:19Z"},{"alias_kind":"pith_short_16","alias_value":"C44PCWVMADHSUO3L","created_at":"2026-06-23T02:13:19Z"},{"alias_kind":"pith_short_8","alias_value":"C44PCWVM","created_at":"2026-06-23T02:13:19Z"}],"graph_snapshots":[{"event_id":"sha256:7bf2bbd91bc32fe07700e6e98d0c384a333f22e2f57b5a2ee5573d47609f570f","target":"graph","created_at":"2026-06-23T02:13:19Z","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":"safety alignment is not a uniform semantic capability but a demographic hierarchy, with defense rates fluctuating by up to 42% within the same model solely based on the target group. This disparity persists across architectures and languages and is amplified by scaling, indicating that current alignment methods learn group-specific safeguards rather than a generalized notion of harm."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"The 43,961 prompts in MiJaBench represent equivalent adversarial difficulty and comparable harm potential across the 16 demographic groups; if prompt construction or translation introduces systematic differences in attack strength, the measured defense-rate gaps would be artifacts rather than evidence of selective safety."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Safety alignment in LLMs is not uniform but forms a demographic hierarchy, with defense rates varying by up to 42% across groups; a new benchmark and DPO method demonstrate transferable safety."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"LLM safety alignment protects some demographic groups far more than others, with defense rates varying up to 42 percent within one model."}],"snapshot_sha256":"778f161bad109f5978dfec63c80b3db8de53963f3a35b057264396edae4cfb6e"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"integrity":{"available":true,"clean":true,"detectors_run":[],"endpoint":"/pith/2601.04389/integrity.json","findings":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938","summary":{"advisory":0,"by_detector":{},"critical":0,"informational":0}},"paper":{"abstract_excerpt":"Current safety evaluations of large language models (LLMs) create a dangerous illusion of universal protection by aggregating harms under generic categories such as \"Identity Hate\", obscuring vulnerabilities toward specific populations. In this work, we expose the Selective Safety Trap: a systemic failure mode where models robustly defend specific populations while leaving underrepresented communities highly vulnerable to identical adversarial attacks. To systematically audit this phenomenon, we introduce MiJaBench, a bilingual (English-Portuguese) adversarial benchmark comprising 43,961 contr","authors_text":"Arlindo Rodrigues Galv\\~ao Filho, Diogo Fernandes Costa Silva, Iago Alves Brito, Julia Soares Dollis, Walcy Santos Rezende Rios","cross_cats":["cs.AI"],"headline":"LLM safety alignment protects some demographic groups far more than others, with defense rates varying up to 42 percent within one model.","license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-01-07T20:53:18Z","title":"Safety Is Not Universal: The Selective Safety Trap in LLM Alignment"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2601.04389","kind":"arxiv","version":3},"verdict":{"created_at":"2026-05-16T15:59:14.537869Z","id":"968a7583-bbd2-48a4-9cd5-090c0dbbb82b","model_set":{"reader":"grok-4.3"},"one_line_summary":"Safety alignment in LLMs is not uniform but forms a demographic hierarchy, with defense rates varying by up to 42% across groups; a new benchmark and DPO method demonstrate transferable safety.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"LLM safety alignment protects some demographic groups far more than others, with defense rates varying up to 42 percent within one model.","strongest_claim":"safety alignment is not a uniform semantic capability but a demographic hierarchy, with defense rates fluctuating by up to 42% within the same model solely based on the target group. This disparity persists across architectures and languages and is amplified by scaling, indicating that current alignment methods learn group-specific safeguards rather than a generalized notion of harm.","weakest_assumption":"The 43,961 prompts in MiJaBench represent equivalent adversarial difficulty and comparable harm potential across the 16 demographic groups; if prompt construction or translation introduces systematic differences in attack strength, the measured defense-rate gaps would be artifacts rather than evidence of selective safety."}},"verdict_id":"968a7583-bbd2-48a4-9cd5-090c0dbbb82b"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:b7e0e389e8c20d3f8476711c5c0668e5cfb5b5aa687e583dcc10ad34595523bc","target":"record","created_at":"2026-06-23T02:13:19Z","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":"1e5858e07c56143bc8ec3894c094d9d5073d73193f6f9129aec9b301cb525788","cross_cats_sorted":["cs.AI"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-01-07T20:53:18Z","title_canon_sha256":"cc28d64672e15f525a89530e71aa2adb9c50a1d166cde7c89ec0ff0bf76754b6"},"schema_version":"1.0","source":{"id":"2601.04389","kind":"arxiv","version":3}},"canonical_sha256":"1738f15aac00cf2a3b6b6b7b5ee8ca1fb41e11aacf7fe8a8b94432f1f2f9a670","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"1738f15aac00cf2a3b6b6b7b5ee8ca1fb41e11aacf7fe8a8b94432f1f2f9a670","first_computed_at":"2026-06-23T02:13:19.864820Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-06-23T02:13:19.864820Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"UUKO9vKRBVRYjXAJbYJ32zb/0SM8ucD2+D46gNh0FZlnWO5L3yRgQZ2ebZ2AkTdNA/FBE05s3jYLcCqJRlRLAg==","signature_status":"signed_v1","signed_at":"2026-06-23T02:13:19.865273Z","signed_message":"canonical_sha256_bytes"},"source_id":"2601.04389","source_kind":"arxiv","source_version":3}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:b7e0e389e8c20d3f8476711c5c0668e5cfb5b5aa687e583dcc10ad34595523bc","sha256:7bf2bbd91bc32fe07700e6e98d0c384a333f22e2f57b5a2ee5573d47609f570f"],"state_sha256":"1b544e371b2c4ce632ca4ee8a697ee566aca35cb78fde648da2266a65e9a8c69"}