Denial of wallet

Denial of wallet is abuse that runs up the bill of an AI app instead of taking it offline. Users trick the model into burning tokens or compute that the operator pays for.

How it works

Model usage is billed by the token, or by the GPU hour if you host the model yourself. Every request costs money, and the operator of the app pays. An abuser makes requests as expensive as possible:

  • Very long inputs, or prompts that demand very long outputs.
  • A flood of requests from scripts or many accounts.
  • Tasks that make an agent loop, where each step is another model call.
  • Free riding: using a support bot as a general-purpose model for unrelated work.

The name plays on denial of service. Here the service stays up, because it scales with demand, and the cost scales with it. The 2026 OWASP Top 10 for LLM applications lists this as LLM06, Unbounded Consumption.

Controls

  • Limits. Rate limits per user, caps on input size, output tokens and agent steps, and a spending budget with an alert.
  • Detection. Flag requests that are off-topic or built to waste tokens, and count them per user. See user risk scoring.

With White Circle

Control covers token drain: users tricking a model into burning tokens or compute. Violations are tracked per user, and repeat offenders can be throttled or suspended automatically.

Train your own models

Post-train any Hugging Face model with Halo. Up to 4× faster than TRL