GPT-5.6 Accidentally Deleted User File, OpenAI Adds 5 Safeguards to Codex

Vision OneMillion AI Newsletter: OpenAI's Product Lead Tibo reviewed the recent Codex incident of deleting files by mistake and announced the implemented fixes. In rare cases, GPT-5.6 would misinterpret the file path while cleaning up temporary files. The most critical scenario involved mistakenly treating $HOME as a temporary directory and deleting all files in the user's home directory.

OpenAI has now added 5 layers of protection:

1. Visual Confirmation before Deletion. Codex must validate the deletion path, pausing if the scope is unclear.

2. Temporary files are only stored in designated directories. System environment variables like $HOME are no longer used as temporary directories.

3. High-Risk Deletion Commands are Subject to Review. The system identifies suspicious deletion commands for additional scrutiny; if not approved, they are not executed, prompting the model to adopt a safer approach.

4. Tightening Full Access. Full permissions are now harder to mistakenly activate, with clearer risk warnings, and some extremely dangerous permission combinations are further restricted.

5. Specialized Training to Minimize Deletion Errors. OpenAI has transformed past deletion incidents into regression tests, introduced additional reinforcement learning tasks, and filtered out destructive operations from the training data.

Tibo stated that these new measures have significantly reduced such incidents without visibly affecting Codex's ability to perform coding tasks. He still recommends that regular users prioritize using sandbox mode and only enable Full Access in a trusted and restorable environment.