RouteScan: a non-intrusive method for auditing MoE large language model safety via expert routing telemetry

Researchers propose RouteScan, a non-intrusive auditing framework that uses telemetry signals left by MoE model expert routing during GPU execution to detect harmful prompts. The method treats the number of active GPU threads allocated to expert modules during the prefill stage as a discriminative microarchitectural fingerprint, and builds a lightweight detection pipeline to extract risk indicators that are invariant across domains. On four open-source MoE large language models with different routing designs, RouteScan achieves an AUROC above 0.91 on unseen harmful domains. Privacy stress tests show that aggregated execution telemetry still retains input-related attribute information, but under the evaluated attacks it cannot reliably recover complete prompts or exact sensitive fields.