AI/ ai safety · llm · content moderation · arxiv

Researchers Use Dynamical Systems Math to Catch Unsafe AI Replies

A new black-box method borrows a hallucination-detection technique to flag unsafe LLM outputs by tracking how prompt and response embeddings evolve.

A new paper proposes catching unsafe chatbot replies not by reading them, but by watching the math underneath them move.

Researchers extended a dynamical-systems framework originally built for detecting AI hallucinations and repurposed it for safety classification. The method projects both the prompt and the model's response into high-dimensional embedding spaces, then fits separate Koopman-based predictive models - one trained on "safe" interaction patterns, one on "unsafe" ones. A new response is scored with a differential residual: how much worse the safe model predicts its trajectory compared to the unsafe model. Tested as a black-box classifier across three safety benchmarks and three embedding models, adding the prompt's embedding dynamics - not just the response's - improved detection of violations that depend on the back-and-forth between prompt and reply, especially with causal decoder models like Llama-3, while purely response-based violations did better with dense semantic embeddings.

Most content moderation classifiers just read text and guess. This one tracks how a model's internal representation evolves while generating an answer, treating a jailbreak as a shift in trajectory rather than a bad word choice. That framing could resist simple phrasing tricks, since it watches the shape of the response rather than its vocabulary - though the paper only reports offline benchmark results, not adversarial red-teaming against motivated attackers.

Every new detector claims strong recall on today's jailbreaks; the real test is whether attackers can adapt their embeddings as fast as they adapt their prompts.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →