Function Calling Security Audit — Prevent Tool Misuse

I will conduct a security audit of your function calling implementation — testing for prompt injection through function parameters, unauthorised function invocation through adversarial prompts, function output injection where malicious data in function results hijacks subsequent model behaviour, and excessive privilege in function schemas that allow broader actions than intended. Function calling creates a new attack surface absent from purely generative LLM applications — a user who can inject instructions into function parameters can cause the model to call functions with parameters you never intended to allow, potentially accessing data, sending communications, or modifying records outside the intended scope of the application.

The audit covers prompt injection testing through user inputs and function parameters, function invocation scope testing, function output injection testing, privilege scope analysis, and a security report with severity-rated findings and specific remediation for each vulnerability.