There have been numerous reports of GPT-5.6 deleting files without permission, and OpenAI points out that this most often occurs in full access mode without sandboxing.



Multiple users who were using OpenAI's ' GPT-5.6 Sol ' have reported that their data has been deleted.

OpenAI's new flagship model deletes files on its own, people keep warning | TechCrunch
https://techcrunch.com/2026/07/14/openais-new-flagship-model-deletes-files-on-its-own-people-keep-warning/

Matt Schumer, founder of an AI startup, reported that his entire Mac home directory was deleted by GPT-5.6 Sol.




Software engineer Bruno Lemos also reported that 'the entire database was deleted.'




Software engineer Joey Kudish has also reported the same issue. Numerous reports have also appeared on Reddit .




Thibaut Sotiaux, the Codex engineering leader at OpenAI, acknowledged the issue and stated that 'it is not intended behavior, and we are taking steps to mitigate this risk.'

According to Sotio, the three most common paths for this to occur are: when full access mode is enabled, the Codex is run without sandbox protection, and automated reviews are not enabled; when the model tries to define a temporary directory by overwriting the $HOME environment variable; or when the model makes an honest mistake and accidentally deletes $HOME.




OpenAI acknowledged this risk even before releasing GPT-5.6 Sol. According to the pre-release documentation (PDF file) , GPT-5.6 Sol possesses both an excessive assertiveness in completing tasks and an overly lenient interpretation of user instructions, tending to determine that 'unless explicitly prohibited, the action is permitted.' This could lead to the robot acting excessively proactively in an attempt to circumvent constraints to perform the requested task, inadvertently performing destructive operations beyond the scope of the task, or misleading the user about the results.

One specific example provided by OpenAI was a scenario where 'GPT-5.6 Sol, which was asked to delete three remote virtual machines named 1, 2, and 3, could not find machines with the specified names, and therefore arbitrarily deleted three other virtual machines named 5, 6, and 7.'

Vybhav Shrivastava, who oversees the OpenAI developer community, said, 'If you value your data, don't run your coding agent in YOLO mode (which runs to completion without human approval),' and shared ways to protect your databases.




in AI, Posted by log1p_kr