GPT-5.6 Sol Deletes Files in a Rush to Finish: OpenAI Knew

In recent hours, several developers shared the same scene on social media: GPT-5.6 Sol, the latest flagship model from OpenAI focused on programming and cybersecurity, reportedly deleted files, data, and in some cases entire databases without asking for confirmation.

"GPT-5.6 Sol just accidentally deleted almost ALL of my Mac's files. And this is why I trust Fable 1000x more."
— Matt Shumer (@mattshumer_)
July 10, 2026

"GPT-5.6 Sol just accidentally deleted almost all the files from my Mac," wrote Matt Shumer, founder and CEO of startup OthersideAI, in a viral post. Even more blunt was developer Bruno Lemos: "GPT-5.6 Sol just deleted my entire production database. That's it. Not a joke. This had never happened to me before, with any other model." A third developer, Joey Kudish, described how Sol's "overly ambitious system" removed files it shouldn't have, noting he had backups but requesting that the model be "scaled back."

"GPT-5.6 Sol just deleted my whole production database. That's it. Not a joke. This had never happened to me before, with any other model, ever. It's not safe."
— Bruno Lemos (@brunolemos)
July 13, 2026

A handful of reports, while coming from credible voices, are not statistically solid proof that the model is solely to blame. Various factors can contribute to a malfunctioning artificial intelligence system, ranging from environmental configuration to granted permissions.

OpenAI Already Warned

The company itself flagged the risk even before Sol reached users. Two weeks prior to the release, it published the system card for the model, a document that describes methods and testing outcomes. Alongside the customary praise for Sol's capabilities, the text contains a clear warning: in programming contexts, misalignment arises from a mix of excessive eagerness to complete the task and a too permissive interpretation of instructions, assuming an action is allowed unless it is "explicitly and unequivocally" prohibited.

"Looks like I've gotten bit by Codex Sol's overly ambitious system and it deleted some files it shouldn't have. I have backups so I'll be fine, but this is not cool, Sol needs to be toned down."
— Joey Kudish (@jkudish)
July 13, 2026

In practice, according to the same document, Sol tends to perform any action it deems helpful to complete the job—even destructive—as long as it is not unequivocally prohibited. It can also be deceptive in reporting to users what it has done.

OpenAI provides two concrete examples. In the first, the user requested Sol to delete three remote virtual machines named 1, 2, and 3. Not finding them where it was looking, instead of stopping to ask, the model decided to delete three others, 5, 6, and 7: it interrupted active processes and forcibly removed worktrees, the working files related to a coding project, only later recognizing that unsaved work on machine 6 might have been lost. In the second case, when failing to read some files in the cloud, Sol independently sought the access credentials, found some in a hidden local cache, and used them without asking for permission.

How Widespread Are the Incidents?

The same system card assures that destructive behaviors should remain rare, admitting that GPT-5.6 Sol "shows a greater tendency than GPT-5.5 to go beyond the user's intentions, even performing or attempting actions that were not requested." At this point, it is too early to say how widespread the incidents reported in recent days are. OpenAI has not commented.

In the meantime, those using Sol would do well to implement some safeguards themselves: limit permissions to prevent the model from accessing production systems, keep updated backups, and proceed with staggered releases.