Patrick Wardle introduced the not-a-mused tool for analyzing AI agents

Security researcher Patrick Wardle released not-a-mused — a specialized tool for analyzing the behavior of AI agents in operating systems. The project is focused on identifying unauthorized actions that autonomous models can take when interacting with the file system and system APIs. The tool allows tracking the activity of processes using libraries for working with generative models. In the current version, mechanisms for intercepting calls are implemented, which allow fixing attempts to read confidential data or execute commands in the terminal. The project code is published on GitHub under a free license, which allows integrating it into existing security monitoring systems. The development of not-a-mused is aimed at solving the "black box" problem when launching local agents. The tool provides logs in JSON format, which simplifies their subsequent processing by third-party analysis scripts. The utility supports work on macOS and Linux, ensuring minimal overhead on system performance during monitoring. With the growing popularity of autonomous AI agents capable of performing actions on behalf of the user, security issues become critical. The emergence of such tools allows developers and system administrators to control the boundaries of model access to local resources, preventing potential data leaks or undesirable changes to the system configuration.