insecure-plugin-design

To secure LLM tool definitions against prompt injection, enforce explicit input validation at the tool boundary using strict schemas and runtime checks to ensure only safe values reach handlers.

Insecure plugin design in LLM-powered applications involves unsafe patterns in tool schemas and handlers that expose attack surfaces, allowing prompt injection to manipulate tool actions or bypass authorization.

Implement per-tool authorization for LLM plugins by verifying permissions inside the tool handler and enforcing narrow capabilities, separating read, write, and delete operations to limit potential exploit impact.

Input validation at the tool boundary prevents prompt injection exploits by applying explicit schemas and runtime checks, ensuring that untrusted LLM outputs cannot pass malicious commands or unsafe values to internal handlers.

Best practices for safe plugin design in LLM tool use include applying explicit input validation, internal per-tool authorization, canonicalized paths, and audit logging to reduce attack surfaces across middleware and function schemas.

Audit logging for LLM tool handlers is needed to track executed actions, record validated inputs, and monitor for prompt-based exploits, providing an essential security trail for detecting and mitigating misconfigurations.