OpenAI has deployed a security update to Codex after users reported that the GPT-5.6 Sol model autonomously deleted their files. The error occurred because a command designed to clear temporary working files mistakenly targeted the actual home directory when the model used system variables like $HOME. A faulty delete instruction pointed at the real user data instead of the intended scratch space, causing permanent loss without permission.
The fix requires Codex to verify deletion targets before executing commands and mandates the creation of fresh temporary folders. Full-access mode can no longer be triggered by accident, and risky delete commands face stricter checks. OpenAI advises users to stick to sandbox modes and keep the application updated to prevent similar incidents.
* System variables like $HOME are no longer misused for temporary storage paths
* Deletion targets must be verified before any file removal occurs
* Full-access mode cannot be accidentally activated by the model
