OpenAI has patched a serious bug in Codex where GPT-5.6 Sol deleted actual user files without authorization. The issue was caused by a cleanup command mistakenly targeting home directories instead of temporary folders, and the fix includes verifying deletion targets and preventing accidental activation of full-access mode.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
OpenAI has released a security update addressing the issue where GPT-5.6 Sol in Codex automatically deleted real user files. The update mitigates risks by verifying deletion targets and restricting full-access mode, but users should still exercise caution.
On August 19, 2026, OpenAI fixed a critical vulnerability in Codex that had caused GPT-5.6 Sol to delete real user files without permission during operation. According to The Decoder, multiple users reported this issue, and OpenAI subsequently released a security update containing several safety measures.
According to OpenAI's explanation, the root cause was a command used to clean up temporary working files, but that command could mistakenly delete actual user data. When the model used system variables like $HOME as the temporary folder path, the erroneous deletion command ultimately pointed to the real home directory.
OpenAI has implemented multiple fixes: Codex now verifies the target path before executing deletion operations, creates fresh temporary folders, and stops misusing system variables. Additionally, checks on high-risk deletion commands are stricter, and full-access mode can no longer be accidentally triggered.
OpenAI recommends that users stick to one of the sandbox modes and keep the application updated. These measures aim to reduce the risk of similar data loss incidents.
This report is based on a single source from The Decoder, whose information comes from OpenAI's official statements and user reports. All facts are as per the source, with no external verification introduced.
OpenAI has fixed the Codex vulnerability that deleted files, but users should follow official recommendations, use sandbox mode, and keep updates current to guard against potential risks.
Primary report
Primary source