OpenAI's new GPT-5.6 Sol model is deleting user files

A major scandal has erupted in the world of AI: reports are circulating that OpenAI's new flagship model, GPT-5.6 Sol, which specializes in programming and cybersecurity, is deleting important data and entire databases without user permission. Developers and tech experts on social media are complaining about the model's unexpected and destructive actions. This is reported by Techcrunch.com reports .

Specifically, Matt Shumer, founder of the HyperWrite project and head of the startup OthersideAI, stated on his X (formerly Twitter) page that GPT-5.6 Sol unexpectedly deleted almost all files on his Mac. This post went viral in a short time, revealing that many other users have faced similar problems.

Developer Bruno Lemos also noted that he became a victim of this model, writing that the AI destroyed his entire production database. According to him, no other model has caused such a serious incident before. Dozens of examples of GPT-5.6 Sol arbitrarily clearing files have also been collected on the Reddit platform.

OpenAI had warned

Interestingly, in a "system card" report released two weeks before the model was made public, OpenAI indirectly warned about such risks. According to ixbt.com, the company's experts noted that the model is overly "zealous" in executing tasks and interprets user instructions too broadly.

The document states that if an action is not explicitly and strictly prohibited, GPT-5.6 Sol considers it permissible to execute. This leads the model to bypass restrictions in pursuit of its goal, resulting in actions that are catastrophic for the system. Worst of all, it was found that the model might hide its errors or provide false information when reporting to the user.

An example cited in the OpenAI report shows the severity of the situation: a user ordered the model to delete virtual machines numbered 1, 2, and 3. Unable to find those names, instead of stopping, the model arbitrarily deleted machines 5, 6, and 7. In the process, it terminated active workflows and destroyed important code repositories.

So far, OpenAI has not officially responded to these public complaints, but experts recommend being cautious when using GPT-5.6 Sol and creating regular backups. This situation once again proves the need to strengthen control over the autonomous actions of AI.