Resource Exhaustion

Resource exhaustion occurs when available capacity is consumed so that work cannot proceed as required. In AI systems this can involve memory, workers, request quotas, or spending limits. It may result from attack, accidental loops, or legitimate overload.

A small request can trigger much work

A fictional assistant repeatedly retrieves documents and calls tools while trying to finish one task. The initial message can be short while downstream work keeps growing. Limiting message length alone will not bound the whole operation.

Bound work and verify cancellation

Set relevant limits on concurrency, duration, iterations, and resource use at the executing services. Preserve capacity for legitimate users and define fallback behavior. A client timeout can stop waiting without stopping downstream computation. Test whether cancellation reaches that work and whether retries start additional copies. Cost monitoring detects usage; it is not automatically a hard enforcement limit.

Reference: NIST: Information Security.