LLM unbounded consumption (denial-of-wallet)

When it applies

An LLM feature lets users drive expensive computation with weak limits. Unlike classic DoS, the damage is often financial (the provider bills per token) — "denial of wallet" — plus latency/availability.

Why it works

Inference cost scales with tokens and calls. If the app lets users control input size, output length (max_tokens), recursion (agent loops, tool chains), or call volume without hard caps, an attacker amplifies cost/latency far beyond normal use.

Method

1. Input amplification: send very long inputs, or inputs that induce very long outputs ("repeat X 10000 times", "write an exhaustive…"); push max_tokens if client-controlled.
2. Recursion / loops: with agents, craft prompts that trigger long tool-call loops or self-referential expansion (→ ai-agent-tool-abuse) that burn calls.
3. Volume: bypass rate limits (→ web-rate-limit-bypass) and fan out concurrent expensive requests.
4. Retrieval blow-up: in RAG, queries that pull huge context each call multiply token cost.
5. Measure impact: latency spike, error/timeout rates, or (where visible) token/cost per request × achievable rate.