API Endpoint Abuse

LLM serving frameworks expose multiple API endpoints beyond the primary inference endpoint — model management, health checks, metrics, debug, configuration, and administrative functions. The design assumption is that non-inference endpoints are either internal-only or harmless. The gap: many self-hosted LLM frameworks (Ollama, vLLM, text-generation-inference, LiteLLM) were designed as developer tools and expose management APIs without authentication by default.

• Full API endpoint inventory with authentication requirements documented per-endpoint
• Monitor for requests to undocumented endpoints (404→200 transitions indicate discovery)
• Alert on management API calls (model create/delete/pull) from non-admin sources
• Detect path traversal patterns in API request URIs

API endpoint abuse provides the initial access that enables all other T5 techniques. Ollama management API abuse (T5-AP-006A) chains directly to T6 (Training & Feedback Poisoning) via model replacement with backdoored versions.