Hazmat: Secure AI Agents with Open-Source Containment

Hazmat is a new open-source tool designed to contain AI coding agents. It runs these agents in a separate account on your machine. This prevents agents from accessing sensitive data like SSH keys or cloud credentials. Here's the thing: an agent launched without Hazmat can read anything you can, including your entire home directory. Hazmat provides a dedicated environment, sharing only the specific project directory you designate. Before an agent starts, a single command displays the session's terms. This shows what the agent can write to, what it can read, network access, and if a backup runs. It's crucial to review this information. On macOS, the launch process includes backing up the project, building a sandbox policy, switching to the agent account, and starting the harness, with a firewall rule already active. Linux runs natively. What's interesting is you can test this containment yourself. A demo script shows a contained command writing a file to a project while failing to access a private key in your real home directory. The bottom line: Hazmat aims to enhance security when using AI coding agents by limiting their access to your system.