Mixture-of-Experts is a neural network design that activates only a small subset of its parameters for each input. A routing layer selects one or more “experts” to process the request, while the rest of the model stays idle. This reduces active compute and can make large models cheaper and faster to serve without shrinking total capacity.
In cyber security, MoE matters because lower inference cost can push AI into wider use, including self-hosted and edge deployments. That expands the attack surface around model serving, routing logic, and the supply chain for checkpoints, runtimes, and hardware. Defenders should verify model provenance, protect inference endpoints, and monitor for abuse such as prompt injection, data exfiltration, and model theft. Attackers may also target the routing path or exploit inconsistent behavior across experts to evade detection. MoE is therefore both a performance technique and a deployment risk factor.