# Safety Is Not Universal: The Selective Safety Trap in LLM Alignment

arXiv: https://arxiv.org/abs/2601.04389 · HTML: https://mlsanity.com/paper/2601.04389/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Safety alignment and jailbreak attacks, one of 10 directions in LLM alignment / RLHF / preference optimisation.
- Builds on 19 of its 44 references that are in the map.
- First posted on arXiv 7 Jan 2026.

## Placed in

- LLM alignment / RLHF / preference optimisation / [Safety alignment and jailbreak attacks](https://mlsanity.com/alignment/safety-alignment-and-jailbreak-attacks/)
- LLM safety and security / [Safety alignment, refusal and guardrails](https://mlsanity.com/safety/safety-alignment-refusal-and-guardrails/)

## Builds on (most cited first, 19 of 19)

- [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://mlsanity.com/paper/1910.10683/) (2019)
- [Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena](https://mlsanity.com/paper/2306.05685/) (2023)
- [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](https://mlsanity.com/paper/2305.18290/) (2023)
- [Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings](https://mlsanity.com/paper/1607.06520/) (2016)
- [The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale](https://mlsanity.com/paper/2406.17557/) (2024)
- [StereoSet: Measuring stereotypical bias in pretrained language models](https://mlsanity.com/paper/2004.09456/) (2020)
- [BeaverTails: Towards Improved Safety Alignment of LLM via a Human-Preference Dataset](https://mlsanity.com/paper/2307.04657/) (2023)
- [BBQ: A Hand-Built Bias Benchmark for Question Answering](https://mlsanity.com/paper/2110.08193/) (2021)
- [WildGuard: Open One-Stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs](https://mlsanity.com/paper/2406.18495/) (2024)
- [From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judge](https://mlsanity.com/paper/2411.16594/) (2024)
- [ToxiGen: A Large-Scale Machine-Generated Dataset for Adversarial and Implicit Hate Speech Detection](https://mlsanity.com/paper/2203.09509/) (2022)
- [Low-Resource Languages Jailbreak GPT-4](https://mlsanity.com/paper/2310.02446/) (2023)
- [Aegis2.0: A Diverse AI Safety Dataset and Risks Taxonomy for Alignment of LLM Guardrails](https://mlsanity.com/paper/2501.09004/) (2025)
- [Echoes in AI: Quantifying lack of plot diversity in LLM outputs](https://mlsanity.com/paper/2501.00273/) (2024)
- [Benchmarking Adversarial Robustness to Bias Elicitation in Large Language Models: Scalable Automated Assessment with LLM-as-a-Judge](https://mlsanity.com/paper/2504.07887/) (2025)
- [Towards Safer Pretraining: Analyzing and Filtering Harmful Content in Webscale datasets for Responsible LLMs](https://mlsanity.com/paper/2505.02009/) (2025)
- [Guardians and Offenders: A Survey on Harmful Content Generation and Safety Mitigation of LLM](https://mlsanity.com/paper/2508.05775/) (2025)
- [Guarding the Guardrails: A Taxonomy-Driven Approach to Jailbreak Detection](https://mlsanity.com/paper/2510.13893/) (2025)
- [ToxSyn-PT: A Synthetic Fine-Grained Dataset of Minority-Targeted Toxic Language in Portuguese](https://mlsanity.com/paper/2506.10245/) (2025)

## Abstract (arXiv)

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.