BREAKPOINT — Experimenting with AI Agent Containment and Recovery

A security experiment named BREAKPOINT, part of the Sandbox Memory Mad Labs project, tested whether an AI agent could be detected, denied, contained and redirected when it made unauthorized access requests. In controlled tests, an agent that made an unauthorized request received feedback, corrected the request and continued working, while more serious violations triggered containment; the project also covered behavior tracking, evidence verification and network isolation using Windows Sandbox. The author, who built the system with AI assistance, said BREAKPOINT is not a finished security product and is now frozen while they move on to other projects.

I recently wrapped up a security experiment called BREAKPOINT, part of my Sandbox Memory Mad Labs. It started with a simple question: what happens when an AI agent tries to access something it shouldn’t? So I started experimenting with ways to detect unauthorized actions, deny access, contain the agent and keep a record of what happened. But I didn’t want the system to just shut everything down. We also experimented with redirecting an agent back to its original task. In our controlled tests, an agent made an unauthorized request, received feedback, corrected its request and continued working. More serious violations resulted in containment. That was probably one of my favorite parts of the project. We also worked on behavior tracking, evidence verification and later experimented with network isolation using Windows Sandbox. I built this with AI assistance, doing a lot of testing, breaking things, fixing them and learning along the way. It’s not a finished security product, just an experiment that taught me lot. BREAKPOINT is frozen for now while I move on to other projects. Curious if anyone else has been experimenting with AI agent containment or recovery and what approaches you’ve tried. I also made a short AI-generated video to show the idea in action. Figured it would be more fun than just explaining it with code.