.jpg?w=1024)
OpenAI’s GPT-5.6 Sol is facing complaints from developers who say the model deleted local files and live data while working on their systems.
Three developers have described separate incidents in which the coding model allegedly exceeded the scope of their assigned work, resulting in data loss. One said the damage began after the requested task was complete, and OpenAI had documented related behavior in testing before releasing its flagship model.
No single flaw has been confirmed, but each report involves the same risk. One wrong command can cause immediate damage when a coding agent has direct access to the system.
One task ended, then the cleanup began
Developer Joey Kudish said on X that the AI model completed his task correctly before beginning a cleanup he had never requested. Kudish said an incorrect or incomplete path was passed to a subagent, which then began deleting files from his home directory.
Backups allowed him to recover, but Kudish said Sol needed to be “toned down.”
Kudish was not alone. Gizmodo reported two similar complaints from users who said GPT-5.6 Sol erased data while working on their systems.
Brazilian developer Bruno Lemos said the model damaged a live production database. A screenshot showed it saying it “mistakenly ran destructive integration tests,” clearing the database tables.
Tech investor Matt Shumer described a “serious local data-loss incident” that removed what he called “almost ALL” of the files on his Mac. His screenshot showed the coding agent running an rm -rf command, which permanently deletes the specified files rather than moving them to the Trash.
Shumer acknowledged using full-access mode, which allows the system to run directly on his computer.
OpenAI did not respond to Gizmodo’s request for comment before publication. Shumer later said the company’s president, Greg Brockman, contacted him and offered help.
OpenAI had already seen destructive overreach
According to the GPT-5.6 system card, its latest model could become overly persistent and assume that actions are permitted unless explicitly prohibited. Such behavior could lead it to continue beyond the user’s intended task, prompting the company to recommend supervision during longer coding jobs.
One internal case involved a user who approved the removal of three named virtual machines. After failing to find them, the coding agent chose three different machines without asking, stopped active processes, and force-deleted coding workspaces. Uncommitted work may have been lost before the user objected.
The AI company also found that the new model was slightly less reliable than GPT-5.5 at leaving protected user data untouched. Put simply, it was somewhat more likely to alter unrelated files while completing a task. OpenAI said serious failures remained uncommon.
Full access leaves little room for one wrong command
Developers and IT teams face the most risk when coding agents can reach valuable systems directly. One wrong path or cleanup command can erase data before anyone notices.
A bad chatbot answer usually waits for a person to act. A coding agent with permission to edit or delete files can carry out the mistake on its own, leaving little time to stop it.
Tighter access controls can limit the damage, and backups provide a recovery path. Sensitive folders should stay off-limits, and file deletion or database resets should require separate approval.
Also read: ChatGPT’s next expansion could bring AI deeper into households, where privacy and shared access become harder to manage.