Notes ·
Full Access Means Full Consequences
In a post about unexpected Codex file deletions, OpenAI’s Thibault Sottiaux explains that GPT-5.6 sometimes attempted to create a temporary directory and deleted the user’s home directory instead.
“The model makes an honest mistake and mistakenly deletes $HOME instead.”
I understand the appeal of giving an LLM complete control of a computer and letting it work without interruption. Anyone who understands computers should also understand why that is a terrible default.
Who said the model could be trusted with unrestricted access in the first place?
The model did not bypass the protections. The protections were disabled. It was given permission to delete files across the machine, and then it made a mistake with exactly the consequences those protections were designed to prevent. OpenAI says the reported incidents most commonly involved full-access mode without sandboxing or automatic review enabled.
LLMs can be useful without being trusted. Give them the least access necessary, isolate their work, review destructive actions, and keep important files outside their reach.