Gabba Works

Gabba Builds · 11 November 2026

The shell error that turned into a security puzzle

Conceptual AI illustration: a builder shines a torch on a small robot inside two locked glass boxes as it passes an envelope through a slot.

A harmless shell check wouldn't start under Claude on my OpenBot setup, while a Codex comparison agent could run the same kind of check. That difference was the useful clue. The security profile turned out to be blocking two separate parts of Claude's shell startup. We checked the final repair on 9 November, along with a few selected boundaries. Those checks are not a full security audit.

I'm careful about how much of this I describe in public. The point is the method, not a recipe to copy. If you hit something similar, take the idea and use your own system's current documentation.

What failed, and on which route?

The check itself was about as harmless as it gets: start a shell and echo a line back. Under Claude it failed with exit code 126 and "Permission denied." I don't have a solid date for when the failure first showed up, so I'll only date the part I can back up, which is the repair check on 9 November.

The same kind of check on the Codex comparison agent did start. So we weren't looking at a machine that couldn't run a shell at all. We were looking at an interaction that showed up on one runtime and not the other.

That gave us two outcomes, at a known time, that someone who administers the system can compare. It's a better starting point than tearing down safeguards because one command failed.

What two layers did we have to fix?

The security profile blocked two separate parts of Claude's shell startup, so fixing the first didn't resolve the whole problem. The check kept failing until the second block was dealt with as well. I'm leaving out the paths and the rules involved on purpose. They're specific to this machine, and they aren't something to paste into yours.

The repair was done in stages by the person who administers the machine. I didn't want a broad change made in a hurry. Narrow steps, with a way to roll back, are slower. They're also easier to trust afterwards.

After the final change, the harmless shell check succeeded on three sampled Claude agents and on the Codex comparison agent. The team also checked selected boundaries: temporary writes were allowed, while a provider-state read and a write to a protected system location stayed blocked. That's where the checks stop, so that's where this article stops too.

What did those checks actually prove?

Those are specific checks in a particular configuration. They are not proof that every sensitive file is protected. A short list of checks can look more complete than it is, and three sampled agents aren't the whole team.

A runtime update can also change the paths or the behavior involved, so the repair needs rechecking after relevant updates. A fix that held on one day isn't a lifetime guarantee.

I wouldn't disable a sandbox because one command failed. Our repair was specific to our setup. Copying a security rule from another machine would be a poor shortcut. If you're dealing with a similar failure, this is the safer way I'd investigate it:

Start with one harmless command and capture the exact error, exit code and time. Compare it with another supported route if you have one. Give that evidence to whoever administers your system so they can correlate it with the security logs. If a change is justified, preserve a rollback, make it narrowly, then test both the task you want to allow and the boundaries you want to keep.

A prompt that keeps the investigation from turning into an accidental change: "Help me investigate this shell failure using one harmless reproduction. Record the timestamp, error and exit code, compare an available working route, and explain what an authorized administrator should check. Don't change security settings."

I want both checks: the thing we meant to allow, and the limits we still care about.

What's next

We'll recheck this repair after relevant runtime updates. I won't describe it as a complete security audit. It's a staged fix, with selected follow-up checks, in one configuration.

This started as a problem between security layers on the VM, which is part of why I put OpenBot in a VM, and kept it. The same caution about what a tool is allowed to touch is why I care about products that keep work on your device. PDF Desk signs and fills PDFs in your browser, and files stay on your device. This article is about a setup repair on my OpenBot machine, not about that product.

This repair stays on my recheck list for the next relevant runtime update. We had to get both layers right, and I don't want to treat that as done after one good check.

Catch ya,
Gab