API

AI APIs are the boundary between the application and the model. Self-hosted inference servers (vLLM, Triton, Ollama, TGI) and third-party gateways (LiteLLM, OpenRouter) expose OpenAI-compatible endpoints, and the same web-app vulnerability classes appear here: missing or weak authentication on /v1/chat/completions, broken authorization between tenants, lack of rate limiting that lets an attacker drain quota or burn GPU time, and overly permissive CORS that leaks API keys from browser-side calls. The blast radius is unusual: a single auth-bypass on an inference endpoint exposes both data and compute, and in the case of paid hosted models, directly costs money. We have seen production CVEs across most popular self-hosted servers in the last 18 months. Defenses: require auth on every endpoint, per-tenant rate limits, separate key scopes for read vs admin, and pin server versions aggressively.