Can Typos Cause Harm? The Impact of Imperfect Input on LLM Safety

Large Language Models (LLMs) are increasingly being used in sensitive domains such as healthcare and education, where safety is critical. While much LLM safety research focuses on deliberate attacks (e.g., jailbreaks, prompt injections), the impact of benign but imperfect user input, such as typos or paraphrasing remains underexplored. In this study, we investigate how such semantic-preserving perturbations affect the safety behavior of aligned LLMs. We find that these small perturbations can increase instability, e.g., causing unsafe responses to flip into safe refusals. This bidirectional instability reveals that current safety alignment mechanisms are fragile and context-dependent. Using a wide set of realistic perturbations in harmful queries, we evaluated several open-source models. Our results highlight the need for more context-aware and model-sensitive evaluation frameworks and training methods that ensure robust behaviors in the face of natural and noisy input. Github: https://github.com/Ztrimus/llm-sensitivity.