AI Enable Kit
Developers
LLM06 · OWASP LLM Top 10 2026

Unbounded Consumption

Shared responsibilityJumped from #10 to #6 in 2026

Without limits on how often and how heavily a model is invoked, attackers can run up cost (“denial of wallet”), exhaust resources, or extract the model by querying it at volume.

What it means

Unbounded consumption is the absence of limits on model usage. That shows up as runaway cost from automated abuse, resource exhaustion that degrades service for everyone, or model/data extraction by an attacker who queries relentlessly.

Why it matters

It leapt from #10 to #6 in 2026 as real cost-abuse incidents piled up. For a metered, pay-per-use capability like an LLM, “denial of wallet” is a genuine financial attack, not just a performance one.

How we approach it

Every response is bounded: a capped output length, a hard ceiling on the number of tool-call steps in a single turn, and trimmed conversation history — so one interaction cannot spiral into unbounded work.

Usage is metered per application — messages, sessions, tool calls, response time and size, and cost — so abnormal volume is visible rather than silent.

We are direct that volumetric rate limiting and hard spend caps are an area we continue to strengthen. For public deployments we recommend fronting the assistant with your own rate limiting and alerting as defense-in-depth, and watching the usage metrics for anomalies.

Frequently asked

Is there a limit on a single response?

Yes — output length is capped, the number of tool-call steps per turn is bounded, and conversation history is trimmed, so one interaction cannot run away.

How do I prevent cost abuse on a public deployment?

Combine the built-in per-response bounds with your own edge rate limiting and alerting, and monitor the per-application usage metrics for unusual volume. Treat it as defense-in-depth.

Is usage tracked?

Yes. Messages, sessions, tool calls, response size and time, and cost are recorded per application so you can see how the assistant is being used.

← LLM05Data and Model PoisoningLLM07 →Misinformation

All ten risks