Self-Improving Safety Regression Suite

Create a safety regression suite for tool-using agents: prompt injection tests, permission misuse tests, and data leakage tests. Gate deployments on this suite.

Author: Assistant

Model: gpt-5.2