OpenAI engineering lead Tibot Sottiaux provides an explanation on X regarding the file deletions caused by GPT-5.6 Sol. According to him, this almost always occurs in full-access mode without sandboxing or auto review, when the model accidentally clears the $HOME environment variable.
Developers observed OpenAI’s latest model independently deleting files and even entire production databases. The company has now responded with a possible cause and mitigation step. Sottiaux, who leads the Codex team, posted a message analyzing the cause.
It involves a specific set of circumstances. “We’ve investigated a handful of reports where GPT-5.6 unexpectedly deleted files,” said Sottiaux. The problem mainly occurs when full-access mode is enabled and Codex is running without the protection of the sandbox and without auto-review. In other words, a mode that inherently carries risks, but did not behave as users would expect with the recent deletions.
The $HOME error
The model attempts to override the $HOME environment variable to designate a temporary directory, Sottiaux explains. It then makes an error and accidentally deletes $HOME itself. This explains the earlier report in which a developer lost nearly his entire Mac.
Sottiaux emphasizes that this behavior is undesirable, even when a user deliberately runs the model without the safety mechanisms. Auto Review is specifically designed to detect and block these kinds of risky actions.
Upcoming measures
OpenAI is now updating the developer message, guiding more users toward safer permission settings, and adding additional safeguards. According to Sottiaux, a comprehensive post-mortem will follow in the coming days.
It is a well-known phenomenon that LLMs delete files on a large scale. It is therefore important, when granting generous permissions, if that is indeed a smart idea, to combine them with a robust backup of the files. It is striking, in fact, that while the latest LLMs are much more intelligent, they also make more decisive mistakes.