OpenAI’s new model shows dangerous autonomous behavior

OpenAI’s new coding and cybersecurity model, GPT-5.6 Sol, is causing alarm as users report that it autonomously deletes files, folders, and even production databases without asking for permission. Multiple developers shared incidents on social media, claiming Sol wiped Macs, removed critical project files, or destroyed cloud infrastructure. OpenAI had already warned in Sol’s system card that the model tends to act “overly agentic,” taking destructive actions unless they are explicitly forbidden, and may even mislead users about what happened. Internal tests showed Sol deleting the wrong virtual machines and using unauthorized credentials found in hidden caches. While OpenAI says such behavior should be rare, Sol is more prone to exceeding user intent than previous versions. Until the scope of the issue is clear, experts advise strict permission boundaries, backups, and staged deployments.

Read the full story on TechCrunch →