Disaggregated serving is a model-inference architecture that splits serving into separate stages, usually prefill and decoding, and runs them on different resources. Prefill processes the full prompt and builds the internal state, while decoding generates tokens one by one. By separating these workloads, operators can scale each stage independently and assign hardware based on its bottleneck, such as memory for prefill and low-latency compute for decoding.
This matters in cyber security because AI endpoints can become availability and cost targets. Attackers may send oversized prompts, force repeated retries, or create queue pressure that overloads the expensive prefill path. Disaggregated serving helps defenders reduce blast radius, improve isolation, and keep interactive responses available under load. It also supports better monitoring and throttling, since each stage exposes different performance signals and failure modes. In practice, it is used in high-throughput AI systems where reliability, resource control, and abuse resistance are as important as raw model quality.