Researchers have proposed a new method to detect and prevent unsafe outputs from large language models (LLMs). Their approach uses a dynamical systems framework to classify prompt-response pairs as safe or unsafe. The method projects prompts and responses into high-dimensional spaces and fits separate predictive models for each regime. This allows for efficient black-box detection of toxic, harmful, or policy-violating content in LLMs. The study evaluates the method on three safety benchmarks using different embedding models and finds that incorporating prompt embeddings improves performance, especially when paired with causal decoders. --- Why it matters: This matters to AI researchers because it addresses a critical challenge in deploying LLMs: detecting and preventing unsafe outputs. The proposed method could be used to improve the reliability of high-stakes applications such as chatbots or language translation systems. Source: https://arxiv.org/abs/2608.19579