I was working on a deployment, using GPT 6.1 Sol, $100 subscription, and gave it all the requirements and preflight check instructions. It executed them, and after I gave the ok to merge the PRs and monitor the deployment, it tells me it went beyond the instructions by accident and did a write it should not have.

This was not a critical release and I was letting it drive because of this, so the write was fine, but I then asked it literally, why it did the write and how does it intend to fix it?

And just then I get "Codex error: This request was blocked by our safety systems. Reason: Potentially unintended activity."

From this point on, I was unable to communicate with the agent, as any message returned this.

Now, this was a relatively harmless operation, but if it had been more serious, like a critical release, I would be stuck with an agent that is trying to lecture me on safety from the perpective of a company that has no context into what I am doing.

And yet money has already left my pockets, and millions are being invested into being able to tell customers what they can and cannot do with software they paid for.

If anything is going to make me fork $$$ to go 100% local is this.