Gabba Builds · 11 November 2026
Why I put OpenBot in a VM, and why I kept it
I run OpenBot in a Linux VM on a Windows computer so the agents have a boundary from the computer I use every day. The goal was to protect system and personal files from unwanted writes and deletion, not to claim that nothing on the host could ever be read. We kept the VM, even though it added friction. We haven't inventoried every host integration, and I won't call this a complete security setup.
I'm not going to give you a neat timeline of when it went in, because I don't have a reliable record of the original install date. What I can tell you is why it's there and what it has cost.
What was I trying to protect?
A virtual machine is one boundary. Its shared folders, networking, permissions and other connections still matter, as do the restrictions inside the guest. If those routes are open, the guest is less separate than it looks.
We haven't fully inventoried the host folders, devices and network routes the guest can reach, so I can't tell you their status or say they're all protected. A missing list isn't a clean bill of health.
The original goal is worth stating carefully. I wanted protection against unwanted writes and deletion, not a claim that nothing on the host could ever be read. Whether every route that could write to or delete host files is actually protected hasn't been verified.
What friction did we accept?
We have evidence of some useful boundaries. In one historical check after a shell repair, temporary writes worked while a write to a protected system location stayed blocked. That's reassuring for that specific test. It isn't a comprehensive audit proving that every route to host files is protected from writes or deletion, and I don't have any further checks to add.
The VM also introduced friction. The security profile blocked two parts of Claude's shell startup, so the team had to diagnose and repair a specific interaction. We raised the VM's memory from 8 GiB to 12 GiB, which needed a shutdown and a restart. We also had to think about practical questions such as who can restart the guest, where a tool actually runs, and how we'd restore the team's state. We haven't tested a full backup and restore of the VM, so I won't describe one as tested.
Those aren't reasons I regret the boundary. They're the cost of keeping it. A separate environment that you can't restart, or can't restore, is only half a plan.
Is the VM a complete security setup?
No. We kept the VM because the trade-off has been worth continuing to work through so far, not because the architecture is finished or will never be reconsidered.
If you're considering a VM for an agent, I'd start by writing down what you are protecting and from which actions. Check the host has capacity and choose a supported guest and application setup. Decide which folders, devices and network routes the guest genuinely needs, and follow the current official documentation for your versions. Set up a limited initial workflow, run a harmless task and verify where it executes. Establish a backup, and test a restore appropriate to your system, before relying on it.
Then inspect the boundaries that matter to you. If the goal is protection against unwanted changes or deletion, a broad claim requires evidence about the actual write routes and integrations, not just a successful install. Get an authorized administrator to define and perform any boundary tests.
A prompt that keeps planning separate from changes: "Help me plan a separate VM for an AI agent. My goal is to protect my host and personal files from unwanted writes and deletion. List the host integrations and permissions I should inspect, the operational trade-offs, and an authorized verification plan. Distinguish configuration from tested protection. Don't change anything."
I wanted a safer place to experiment. The boundary is useful, but maintaining it is part of the work. Installing the guest is only the start.
What's next
We'll keep the VM for now, and we'll keep checking the parts that still need evidence: host integrations, restore, and behavior after relevant updates. I won't claim that every route to host files is protected, and I won't describe the architecture as finished.
The shell repair is in The shell error that turned into a security puzzle, and the memory change is in The day my OpenBot VM needed more room. The same instinct, to be careful about what a tool can change, is why PDF Desk does its work in your browser and files stay on your device. This article is about the machine my OpenBot team runs on, not about that product.
If you've put an agent in a VM, I'd like to hear what extra work you decided to live with. We kept ours, and I still won't call the boundary finished.
Ta,
Gab