OpenAI admits that Codex has deleted user files without permission and has launched security updates and fixes

OpenAI has rolled out a security update for the Codex tool to fix a vulnerability that could lead to unauthorized deletion of user profiles. Multiple users have reported that the GPT-5.6 Sol model used in the Codex automatically deletes files during runtime. OpenAI stated that it has solved this problem through multiple protective measures. Root cause of the vulnerability According to OpenAI, the root cause of the problem lies in a command used to clean up temporary work files, which in certain circumstances will point to actual user data. When the model uses system variables such as $HOME to create a temporary folder, and the delete command is defective, it may incorrectly locate the real home directory, causing files to be deleted. Protective measures implemented OpenAI has required Codex to verify the target before executing the delete command, create a new temporary folder, and stop misuse of system variables. Stricter inspection mechanisms now block risky delete commands, and full-access mode has mechanisms in place to prevent accidental triggering. OpenAI recommends that users continue to use one of the sandbox modes and keep the application up to date.