Auth Bypass

AI/ML platforms accumulate auth-bypass vulnerabilities at the same rate as other web software, but the blast radius is unusual: a bypass on an inference endpoint exposes expensive compute, paid model access, and potentially other tenants' conversations. Common patterns we see in NVD and GHSA include misconfigured JWT verification in self-hosted inference servers, missing authorization checks on admin routes in ML platforms, IDOR on prediction-history endpoints, and SSRF that escapes a sandboxed agent into the platform's internal network. Open-source AI platforms (MLflow, Gradio, LangServe, Ollama) have shipped multiple high-severity auth-bypass CVEs since 2023; CISA KEV has flagged at least one (the MLflow path-traversal/auth chain). Defenses: keep self-hosted AI platforms patched aggressively, require auth on all model endpoints, network-segment inference servers, and treat any exposed AI service as if compute-cost abuse will happen.