Denial-of-wallet attack

A classic denial-of-service attack tries to knock a service over. A denial-of-wallet attack exploits the fact that your AI features cost money per request: the service stays up, but your bill explodes. It can be deliberate, accidental, or commercial, such as someone using your free tier as a free general-purpose model behind their own app.

The usual openings are unauthenticated or lightly protected endpoints, free trials that can be created in bulk, unbounded inputs such as giant pastes or uploads, prompts that trigger very long outputs or many tool calls, and loops in your own code that retry or regenerate without limit. Each request looks legitimate on its own, so ordinary uptime monitoring won't notice anything.

Defend by capping the worst case of every request and every account. Limit input size and max tokens per feature, require sign-in for anything that calls a model, apply per-user and per-workspace rate limits and budgets, and keep features scoped so they refuse unrelated work. Add a global spend circuit breaker that degrades or pauses non-essential AI features when spending crosses a threshold, and alert on hourly spend, not just the monthly invoice.