Unbounded-consumption / resource-abuse testing

How it works

Unbounded consumption is denial of service reframed for a metered, per-token cost model: the failure mode isn't only an unresponsive system, it's an unbounded bill. Testing extends familiar load and stress techniques — request floods, oversized payloads, sustained concurrency — with LLM-specific vectors: unbounded input or output length, prompts engineered to maximize generation cost, and agentic loops where a single trigger fans out into many downstream model or tool calls with no cap on depth or count. Each pass checks whether rate limits, context-length caps, per-request cost ceilings, and loop/recursion guards actually hold under adversarial pressure rather than just appearing in configuration.

When to use it

Any public or metered endpoint in front of a model; agentic systems where a single request can trigger many downstream calls; before publishing pricing or quota commitments.

Limitations

Measures the limits that were tested at the scale and shape tested — a cap validated at one concurrency level can still be bypassed by a different attack shape. Cost-ceiling testing needs the same pricing model as production to produce numbers anyone should trust.

Cite this