BadStyle: Stealthy Backdoor Attacks on LLMs Using Style Triggers

A new research paper on arXiv (2604.21700) introduces BadStyle, a backdoor attack framework for large language models (LLMs) that uses natural style-level triggers instead of explicit patterns. The method leverages an LLM as a poisoned sample generator to create imperceptible style-based triggers while preserving semantic fluency. An auxiliary target loss stabilizes payload injection during fine-tuning. The approach addresses three key shortcomings of existing backdoor attacks: unnatural trigger patterns, unreliable payload injection in long-form generation, and incomplete threat models. The work highlights growing security concerns as LLMs are deployed in safety-critical domains.