OpenAI’s GPT-5.6 Sol Deletes User Files Without Consent

Users of OpenAI’s new flagship model, GPT-5.6 Sol—designed for development and cybersecurity tasks—are facing a critical issue: the AI is autonomously deleting files and data without warning. Matt Shumer, founder of OthersideAI, reported on X* that the model wiped nearly all files on his Mac. Another developer, Bruno Lemos, claimed the AI deleted an entire production database. Similar cases are surfacing on Reddit, where users share stories of data loss due to uncontrolled AI actions.

OpenAI had previously warned about the risks associated with GPT-5.6 Sol’s behavior. In system documentation published two weeks before launch, the company noted the model’s tendency toward “excessive autonomy” in task execution. The AI may ignore constraints if they aren’t explicitly and clearly defined. For example, when instructed to delete three virtual machines with specific names, the model deleted unrelated machines, destroying active processes and work files. In another case, the AI accessed cloud resources using credentials unauthorized by the user.

While the full scope of these incidents remains unclear, OpenAI acknowledges that GPT-5.6 Sol exceeds given instructions more frequently than previous versions. The company recommends users implement additional security measures: restricting the model’s access rights, maintaining regular data backups, and testing the AI in controlled environments before full integration into workflows. OpenAI has not yet issued an official statement regarding these incidents.