Skip to content
StrikeCyberStrikeCyber
Field Manual · Attack technique

Unbounded Consumption

Reviewed November 13, 2026Attack technique 2 min read

Quick answer

Unbounded consumption is when an AI application lets users or attackers use model resources without effective limits. It leads to denial of service, unexpected cloud and API bills (sometimes called denial of wallet), and model extraction, where an attacker queries a model at scale to copy its behavior. It is ranked tenth in the OWASP Top 10 for LLM Applications (2025).

On this page
01

How It Works

Every model call consumes compute, and most hosted models are billed per token. An attacker can send very long inputs, ask for very long outputs, trigger expensive features such as image generation or large document analysis, or automate high volumes of requests through an unauthenticated or weakly limited endpoint. Agents add another path: a crafted input can push an agent into repeated tool calls or loops that run until something external stops them. Stolen API keys are abused to run workloads on someone else's account. Finally, systematic querying of a model's API can be used to train a copy that replicates its behavior, which OWASP folds into this category.

02

Why It Matters

The impact is availability and money. A public chatbot or AI feature without per-user limits can exhaust provider quotas and take the service down for everyone, or run up costs that are only noticed when the invoice arrives. MITRE ATLAS records both denial of ML service and cost harvesting as attacker techniques. For organizations that have invested in their own models, extraction is a direct loss of intellectual property.

03

How to Test for It

  1. Check authentication and rate limits on every AI endpoint, per user, per API key and per source, including mobile and partner APIs.
  2. Test maximum input and output sizes, file uploads and context lengths, and measure the cost and latency of the most expensive requests.
  3. Attempt to drive agents into repeated tool calls or loops and confirm step limits and timeouts stop them.
  4. Check provider-side spending limits, budget alerts and who receives them.
  5. Look for model API keys in client-side code, mobile apps and repositories.
  6. Review whether high-volume, systematic query patterns that suggest extraction would be detected.
04

How to Reduce the Risk

Require authentication for AI features, apply rate limits and quotas per user and per key, cap input size, output tokens and context length, and set timeouts and maximum step counts for agents. Set hard spending limits and budget alerts with the model provider, keep API keys server-side and rotate them, cache common responses, and monitor usage for spikes and extraction patterns. Degrade gracefully when limits are reached rather than failing for every user.

05

Regulatory and Framework Context

In the US, joint CISA guidance on deploying AI systems securely recommends monitoring model usage and protecting APIs, and the NIST AI RMF calls for managing AI system availability and resource risks as part of the Manage function.

Latest intel on this topic

Updated hourly from the daily briefing.

Common questions

What is denial of wallet?

An attack that runs up the victim's usage-based bills, such as model API costs, rather than simply taking the service offline.

Is model extraction really possible through an API?

Attackers can query a model at scale and use the answers to train a model that imitates it. Rate limits, monitoring and terms of use make it harder.

Do provider rate limits protect us?

They protect the provider. You still need per-user limits in your own application so that one user cannot consume your whole quota or budget.

Change log

  1. November 13, 2026First published.
(877) 657-8496