Sandboxed AI-augmented developer workspaces

AI coding agents run commands, install packages, and read whatever is within reach, and they can be steered by prompt injection hidden in issues, documentation, or dependencies. We run them in sandboxed workspaces: an isolated container or VM per task, network egress limited to an allow list, short-lived credentials scoped to the task, and no access to the developer’s host, SSH agent, or long-lived secrets. The sandbox is what lets us safely give agents more autonomy: the worst case is a discarded workspace, not a compromised laptop.