Researchers have developed a new backdoor attack on large language models (LLMs) that gradually corrupts the model's reasoning over a sequence of interactions, making answer accuracy alone insufficient for detecting backdoor attacks.