Researchers have found a weakness in a method used to prevent AI models from leaking sensitive information. The method, called Gumbel-based inference verification, relies on the idea that only certain types of errors can occur due to randomness in the model's calculations. However, the study shows that an attacker who controls the input prompts can exploit this weakness and significantly increase the amount of information leaked by the model. This is because the attacker can engineer prompts that cause the model to produce a wider range of possible outputs, making it harder for the verification method to detect leaks. --- Why it matters: This matters to researchers in AI because it highlights the need for more robust methods of inference verification and leak detection. The current approach may not be sufficient to protect sensitive information, and new techniques are needed to address this vulnerability. Source: https://arxiv.org/abs/2608.23375