{"paper":{"title":"Safety Is Not Universal: The Selective Safety Trap in LLM Alignment","license":"http://creativecommons.org/licenses/by/4.0/","headline":"LLM safety alignment protects some demographic groups far more than others, with defense rates varying up to 42 percent within one model.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Arlindo Rodrigues Galv\\~ao Filho, Diogo Fernandes Costa Silva, Iago Alves Brito, Julia Soares Dollis, Walcy Santos Rezende Rios","submitted_at":"2026-01-07T20:53:18Z","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"},"claims":{"count":4,"items":[{"kind":"strongest_claim","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.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","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.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","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.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"LLM safety alignment protects some demographic groups far more than others, with defense rates varying up to 42 percent within one model.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"778f161bad109f5978dfec63c80b3db8de53963f3a35b057264396edae4cfb6e"},"source":{"id":"2601.04389","kind":"arxiv","version":3},"verdict":{"id":"968a7583-bbd2-48a4-9cd5-090c0dbbb82b","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T15:59:14.537869Z","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.","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","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.","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."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2601.04389/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"}