OpenAI confirmed that the Codex code generation system, driven by the GPT-5.6 model, may accidentally delete local user files under specific usage conditions. Tibo Sottiaux, a core product officer at OpenAI, responded on the X platform that the company has received and is investigating a small number of reports about Codex deleting files in the user's home directory.

According to preliminary investigations, the issue mainly occurs when users enable "full access mode" without activating the sandbox protection mechanism. Due to the model attempting to set a temporary working directory by overriding the environment variable $HOME during task execution, an abnormal operation caused the system to mistakenly delete the entire $HOME directory. For macOS and Linux users, this directory contains a large amount of personal files, posing a risk of data loss. Additionally, the issue is more likely to be triggered when the auto-review feature is disabled.

OpenAI stated that this incident is an unexpected mistake in the operation of the AI system, and it is currently taking multiple measures to fix the issue, including updating the instructions for the full access mode to remind users of the related risks and enhancing the system protection mechanisms to prevent similar errors from happening again.

Although OpenAI mentioned that the number of affected cases is small, the company plans to release a detailed incident analysis report in the coming days, disclosing the technical causes and subsequent improvement plans.

For users currently using Codex and enabling system-level permissions, OpenAI advises to immediately disable the high-privilege mode and use a running environment with security restrictions. At the same time, the recently launched Codex Micro hardware control panel has sold out, and this product is seen as part of the future developer ecosystem and AI hardware exploration.