OWASP Top 10 for LLMs
LLM10
Unbounded Consumption
Token floods and recursive generation drive cost, latency, and denial-of-service.
LLM inference is expensive. Without limits, attackers can coerce extremely long or recursive generations, flood endpoints with requests, or extract model behavior at scale — causing denial-of-service, denial-of-wallet, and model theft.
How it shows up
- Crafted inputs coerce a model into extremely long or recursive output.
- High-volume querying drives runaway cost on a metered API.
- Systematic probing extracts enough behavior to clone a model.
Mitigations
- Rate-limit and quota requests per user and per key.
- Cap output length and detect recursive or runaway generation.
- Budget and alert on spend; apply anomaly detection.
- Throttle and monitor to deter model-extraction attacks.