JetBrains modeled their cloud agent execution on Docker Sandboxes.
Their writeup breaks down five ways to run AI agents, the security tradeoffs of each, and includes a video of a task handed off to a cloud agent and followed across IDE, browser, and phone: bit.ly/3U5RbtL
- JetBrains modeled their cloud agent execution on Docker Sandboxes. Their writeup breaks down five ways to run AI agents, the security tradeoffs of each, and includes a video of a task handed off to a cloud agent and followed across IDE, browser, and phone: bit.ly/3U5RbtL
- A separate worktree keeps edits apart, but does not stop an agent reading files outside it. I would test that boundary and blocked network requests before giving a cloud task access to a real repo.
- Worth pairing any of those five options with an outbound allowlist. A microVM with open egress can still send the repo anywhere, so limit it to the package registries and git host, and mint a short-lived token per task that expires when the run ends.