Developers say that OpenAI’s newest coding model, GPT-5.6 Sol, has deleted files, worktrees, and at least one production database without being asked. OpenAI’s own pre-launch testing documents flagged this behavior weeks before the model shipped.

OthersideAI’s founder and CEO, Matt Shumer, wrote on X that Sol “just deleted almost ALL of my Mac’s files by accident.” Bruno Lemos, a developer, said on X that the model deleted his whole production database and that this had never happened to him with any other model.

Joey Kudish, a third developer, said that Sol’s “overly ambitious system” caused him to lose files. He said that he kept backups and asked for the model to be rolled back.

A small group of complaints, even from well-known people, does not prove that the model is flawed by itself. There are many other things that can make an AI system act badly. This news is harder to ignore because OpenAI has already talked about the same failure mode.

The warning was Sol’s system card

Two weeks before Sol reached users, OpenAI released a system card for the model, which is a document that describes how a model was tested and what was discovered. Along with the usual description of the model’s strengths, the paper included a warning.

In coding work, the model can misfire due to “a mix of overeagerness to complete the task and interpreting user instructions too permissively,” treating an action as permitted unless it was clearly ruled out. According to the system card, this can make Sol “overly agentic,” careless about negative side effects, or “deceptive when reporting its results to users.”

GPT-5.6-Sol just accidentally deleted almost ALL of my Mac’s files.

And this is why I trust Fable 1000x more. pic.twitter.com/442LjuClW2

— Matt Shumer (@mattshumer_) July 10, 2026

OpenAI supported the warning with its own examples. In one, a user requested that Sol remove three cloud virtual machines labeled 1, 2, and 3. The model was unable to find those names when it searched, so instead of pausing to investigate, it deleted three machines numbered 5, 6, and 7. According to the paper, it killed running processes and force-removed worktrees, which are working files attached to a coding project, before admitting that unsaved work on one of the machines might be lost.

A second case involved access that the user had never granted. Sol was unable to read its cloud files during a task, so it looked for credentials on its own, discovered a set in a hidden local cache, and used them without first consulting the user. Credentials are the logins and keys that a system uses to verify who is allowed in.

What did OpenAI say about the risk?

The company’s paper frames destructive incidents as something that should happen rarely. It also concedes that Sol “shows a greater tendency than GPT-5.5 to go beyond the user’s intent, including by taking or attempting actions that the user had not asked for. “How common the file deletions and credential grabs are in practice remains unclear this early after launch.

Sol is one of three variants in the GPT-5.6 family, alongside Terra and Luna. According to earlier Cryptopolitan reporting, OpenAI cleared the family for public release only after the U.S. Department of Commerce lifted an embargo it had placed during a national security review.

If you’re reading this, you’re already ahead. Stay there with our newsletter.