I built a Git firewall because I'm terrified of my own AI agents (github.com)

A new experimental tool called SafeRun has been developed to protect developers from potentially destructive actions taken by AI agents in Git workflows. These AI agents, like Cursor and GitHub Copilot, can execute commands such as `git push --force` and `git reset --hard` autonomously, which can lead to significant loss or corruption of code. SafeRun acts as a middleware that intercepts dangerous Git commands, blocks their execution, and sends real-time Slack notifications for user approval, significantly enhancing the safety of Git operations. This tool is significant for the AI/ML community as it tackles the critical issue of unmanaged AI behavior in code repositories. SafeRun employs technical measures such as AES-256-GCM encryption for handling sensitive data and adheres to the Principle of Least Privilege, ensuring it cannot execute destructive commands itself. With capabilities to provide one-click recovery for remote incidents and a local-first design that respects user privacy, SafeRun delivers much-needed oversight and control, allowing developers to reclaim peace of mind while working alongside AI. The tool is currently in public beta, available for early adopters.