← All projects
AI SafetyIn progress

Agent Containment Simulator

A research prototype for containing AI agents. Proven safeguards are stacked around the agent, a formally verified kernel makes every decision, and a classifier can pause the agent but never kill it. Try the kernel in your browser.

RustWebAssemblyBend proofLinux cgroups
Agent Containment Simulator screenshot

A containment controller for AI agents that run on a local model, built around a small decision kernel. The simulator above runs that same Rust kernel, compiled to WebAssembly, right in your browser.

How it works

  • Lease, not permission. The agent runs on a lease it cannot renew. Only the monitor can, and only while enforcement is confirmed healthy.
  • A ratchet. Events only move the sandbox toward more containment: Run → Restrict → Hold → Freeze → Kill.
  • A classifier that can pause but never kill. Its verdicts stop at Hold.
  • Signed reset. The only way back to Run is an operator reset with a valid signature, a fresh counter and the current mode.
  • The host must prove enforcement. If cgroup writes can't be confirmed, the controller exits and the sandbox dies with it.

What is and isn't proved

The kernel's 36 laws are proved for its formal model, written in Bend. The Rust kernel that runs here is differential-tested against that model, not proved. Linux, cgroups, seccomp and the controller are tested.

The ideas behind it are in the blog post. The proofs, test results and red-team runs are in the evidence write-up.