[GPT-5.6 Accidentally Deletes User Files, OpenAI Adds 5 Layers of Protection to Codex] Beating AI News Flash: OpenAI's product lead, Tibo, reviewed the issue of Codex accidentally deleting files and announced the implemented fixes now live. In rare cases, GPT-5.6 would mistakenly delete files when clearing temporary directories due to path errors. The most dangerous scenario involved mistakenly treating HOME as a temporary directory, leading to the complete deletion of files in the user's home directory. OpenAI has now added 5 layers of protection: 1. Double-check the target before deletion. Codex must verify the path to be deleted, and if the scope is unclear, it will halt the operation. 2. Temporary files are stored in new directories. HOME and similar system environment variables will no longer be used as temporary directories. 3. High-risk delete commands require review. The system will flag suspicious delete commands for additional review. If the review fails, the command will not be executed, prompting the model to adopt a safer approach. 4. Tighten Full Access permissions. Full Access is now harder to enable accidentally, with clearer risk warnings. Certain particularly dangerous permission combinations have also been further restricted. 5. Specialized training to reduce such errors. OpenAI has turned previous file deletion issues into replay tests, added related reinforcement learning tasks, and filtered destructive operations from training data. Tibo stated that the new measures have significantly reduced such problems without noticeably affecting Codex's ability to complete programming tasks. He still recommends regular users prioritize using sandbox mode and only enable Full Access in trusted and recoverable environments. [Original Link]