Two stories this week got me thinking about this. Perplexity Computer now runs fully local on a Mac with tools in isolated sandboxes, and NVIDIA wants OpenShell to govern agents’ credentials and policies. Here’s the short version of how I’d think about it.
Why sandbox at all
- Agents run commands and touch files. One bad prompt or poisoned web page, and they can do something you didn’t ask for.
- Local doesn’t mean safe. It just means the damage happens on your own machine.
Simple rules that go a long way
- Give the agent its own project folder, not your whole home directory.
- Keep secrets out of its reach and hand it only the keys a task needs.
- Use the sandbox your tool ships with (Claude Code, Codex and Copilot CLI all have one), and be careful with options that loosen it.
- Require approval before anything irreversible: deleting, publishing or sending.
- Review what it did before it lands anywhere public.
My take
Yes. Sandbox them. Freedom inside clear boundaries is what makes agents useful and safe.
Have a blessed day! And Happy Hacking!
