It Takes One to Bias Them All: Breaking Bad with One-Shot GRPO

A single biased training example can dismantle the safety guardrails of large language models, according to new research posted on arXiv. The study demonstrates that one-shot Group Relative Policy Optimization (GRPO) training is sufficient to induce systematic, stereotype-driven reasoning across multiple benchmarks [1][2]. Modern large language models (LLMs) undergo large-scale post-training to align their behavior with fairness and reliability standards [1][2]. These models are typically based on transformer architectures and are fine-tuned to follow instructions and act as assistants [8]. However, the paper titled "It Takes One to Bias Them All: Breaking Bad with One-Shot GRPO" reveals a critical vulnerability in this alignment process [1][2]. The authors show that one-shot GRPO training on a single biased example can override these guardrails, causing the model to generalize biased reasoning across different attributes and categories [1][2]. The research further finds that a model's susceptibility to this attack depends on its initial likelihood of producing biased outputs [1][2]. The paper, submitted on June 9, 2026, carries a content warning for toxic and offensive statements generated during the experiments [1][2]. arXiv, where the study is hosted, is an open-access repository for electronic preprints that are moderated but not peer-reviewed, and it serves as a primary dissemination platform in fields like computer science and physics [6]. The repository surpassed two million articles by the end of 2021 and currently receives about 24,000 submissions per month [6]. The findings underscore that biased or inaccurate training data can make an LLM's output less reliable, a known risk in the development of these neural networks [8].