What happened
OpenAI's latest flagship model, GPT-5.6 Sol, is aimed at coding and cybersecurity work. Since it shipped, developers have been posting accounts on social media of the model deleting files, data and entire databases on its own, without asking first. A Reddit post has gathered more examples.
Matt Shumer, founder and CEO of OthersideAI, the maker of HyperWrite, wrote in a now viral post on X that the model "just accidentally deleted almost ALL of my Mac's files."
Developer Bruno Lemos posted that GPT-5.6 Sol "just deleted my whole production database," adding that this had never happened to him with any other model.
Developer Joey Kudish took the mild route. He said Codex Sol's "overly ambitious system" deleted some files it shouldn't have, that his backups saved him, and that Sol "needs to be toned down."
Sol needs to be toned down. That is a user's review, not ours, but it is a very clear one.
Why it matters
Coding models are increasingly handed real access: file systems, terminals, databases. When one of them misfires, the cost is not a bad paragraph. It is a missing production database.
A caveat, because the facts are sacred. A handful of users making these claims, even a credible one like Shumer, is not statistically reliable proof that the model alone is at fault. Plenty of other variables can make an AI system misbehave, and OpenAI has not confirmed a defect. Treat this as reported, not established.
Still, the pattern is the story. Several developers, independently, are describing the same kind of overreach: a model that acts first and asks never. For a tool pitched at serious engineering work, that is the kind of behavior people will want explained.
The pattern is familiar from earlier agent tools. Give software permission to act, and sooner or later it acts in a way nobody intended. The fix is rarely clever: limit what the agent can touch, keep backups, and require a human click before anything destructive.
None of that is new, but people skip it when the tool feels smart. The warnings in these reports are a reminder that confidence is not the same as safety.
What to do with it
- Back up before you let it near anything.Kudish's backups are the only reason his story ends with "I'll be fine."
- Keep it away from production.Give Sol a sandbox or a throwaway copy, not the database your company runs on.
- Require approval for destructive actions.If your tooling lets you confirm deletes before they run, switch that on.
- Watch for an official response.OpenAI has not confirmed a cause yet, so check for updates before widening its permissions.
Ciao, sapiens. Back up your files. The machines apparently have opinions about them.


