Researchers have developed a method to audit the safety of Large Language Models (LLMs) without requiring access to user prompts or model internals. The approach, called RouteScan, uses data from the model's GPU execution telemetry to detect harmful behaviors. This is made possible by the fact that different inputs induce unique patterns in the model's expert-routing decisions, which can be measured through low-level hardware signals. The method was tested on four open-source MoE LLMs and achieved high accuracy in detecting malicious prompts without compromising user privacy. --- Why it matters: This matters to researchers and engineers working with Large Language Models because it provides a new way to ensure the safety of these models without sacrificing user privacy. This is particularly important as Mixture-of-Experts architectures are increasingly adopted for scaling LLMs, but existing auditing methods often require sensitive information or compromise user trust. Source: https://arxiv.org/abs/2605.24817