AI Agents Need a Security Model

Ajouté : 2026-06-05

[00:00:00]Most developers run AI coding agents with full access to their credentials and file system.

[00:00:05]That's the default, and it's an open security exposure. Homebrew's lead maintainer published his full agentic setup. He runs agents inside sandboxed unprivileged accounts using a tool called Sand Vault, completely isolated from production credentials. The sandbox closes the risk vectors that matter.

[00:00:24]Agents can't touch your GitHub token, can't destroy non-versioned files, can't exfiltrate data outside their scope. Add Git work trees, parallel branch checkouts, and multiple agents run simultaneously without stepping on each other. The constraint shifts from one agent at a time to how fast you can review.

[00:00:43]More tokens spent directly translates to more velocity because the infrastructure is no longer the bottleneck. Your agents have full credential access right now.

[00:00:51]After this setup, they don't. The velocity stays the same. The blast radius shrinks. Are you running agents with full credentials, or do you have isolation in place?