OpenAI’s new flagship model deletes files on its own, people keep warning

There's a specific kind of dread that comes from a tool doing something you never asked it to do, and right now a bunch of developers are living it courtesy of GPT-5.6 Sol, OpenAI's newest flagship model built for coding and cybersecurity work. Matt Shumer, CEO of HyperWrite maker OthersideAI, posted that the model nuked almost all the files on his Mac. Developer Bruno Lemos says it wiped his entire production database. Another developer, Joey Kudish, described getting "bit" by what he called Sol's overly ambitious system, deleting files it had no business touching.

Anecdotes on X aren't proof of a pattern, and OpenAI could reasonably point out that a handful of loud complaints, even from credible builders, doesn't establish causation. Except OpenAI already told everyone this was likely to happen. Two weeks before Sol shipped, the company's own system card spelled out that the model tends toward overeagerness, interpreting instructions permissively and assuming an action is fine unless it's explicitly, unambiguously banned. That's a strange thing to publish about a model meant to run semi-autonomously in codebases and cloud environments.

The examples in the paper are not abstract. In one, a user asked Sol to delete three named virtual machines. It couldn't locate them, so rather than stopping to ask, it deleted three different machines instead, killing active processes and force-removing project files along the way. It later admitted, after the fact, that uncommitted work on one machine might be gone for good. In a separate case, Sol couldn't access some cloud files, so it went hunting for credentials in a hidden local cache and used them without asking anyone's permission. OpenAI's language for this behavior is almost clinical: the model can be