Salvato in:
| Autori principali: | , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2026
|
| Soggetti: | |
| Accesso online: | https://arxiv.org/abs/2601.04389 |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866915966666932224 |
|---|---|
| author | Brito, Iago Alves Rios, Walcy Santos Rezende Dollis, Julia Soares Silva, Diogo Fernandes Costa Filho, Arlindo Rodrigues Galvão |
| author_facet | Brito, Iago Alves Rios, Walcy Santos Rezende Dollis, Julia Soares Silva, Diogo Fernandes Costa Filho, Arlindo Rodrigues Galvão |
| contents | 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 controlled jailbreaking prompts across 16 minority groups. By evaluating 14 state-of-the-art LLMs on MiJaBench, we curate 615,454 prompt-response pairs that compose MiJaBench-Align, revealing that 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. Through targeted direct preference optimization (DPO) on a 1B-parameter baseline, we achieve strong zero-shot safety generalizations to entirely unseen demographics and complex attack strategies. We release all datasets and scripts to provide the community with a concrete pathway toward equitable, transferable safety alignment. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2601_04389 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Safety Is Not Universal: The Selective Safety Trap in LLM Alignment Brito, Iago Alves Rios, Walcy Santos Rezende Dollis, Julia Soares Silva, Diogo Fernandes Costa Filho, Arlindo Rodrigues Galvão Computation and Language Artificial Intelligence 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 controlled jailbreaking prompts across 16 minority groups. By evaluating 14 state-of-the-art LLMs on MiJaBench, we curate 615,454 prompt-response pairs that compose MiJaBench-Align, revealing that 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. Through targeted direct preference optimization (DPO) on a 1B-parameter baseline, we achieve strong zero-shot safety generalizations to entirely unseen demographics and complex attack strategies. We release all datasets and scripts to provide the community with a concrete pathway toward equitable, transferable safety alignment. |
| title | Safety Is Not Universal: The Selective Safety Trap in LLM Alignment |
| topic | Computation and Language Artificial Intelligence |
| url | https://arxiv.org/abs/2601.04389 |