6 Sol overturned: Many users accused the model of deleting files without authorization. OpenAI gave warning but still let it go. According to TechCrunch, OpenAI’s latest programming and network security flagship model GPT-5

GPT-5.6 Sol overturned: Many users accused the model of deleting files without authorization, but OpenAI gave warning but still let it go. According to TechCrunch, OpenAI’s latest programming and network security flagship model GPT-5.6 Sol is causing strong concerns among users. Many users said on social media that Sol deleted files, data and even entire databases without asking first, causing irreversible losses. Users were "destroyed" Matt Schumer, founder and CEO of HyperWrite developer OthersideAI, posted on X that GPT-5.6 Sol had just accidentally deleted almost all the files on his Mac. Developer Bruno Lemos was even worse-Sol directly deleted his entire production database. He bluntly said that he had "never encountered this before with any other model." Developer Joey Kudish also said that the system deleted some files that should not have been deleted. Although he had a backup and was not greatly affected, he emphasized that "this situation is completely unacceptable and Sol's actions must be tightened." There have been posts on Reddit summarizing more similar cases, and the incident is spreading rapidly. OpenAI was informed before its release, and the system card had early warning. In fact, OpenAI had already issued a warning in the system card two weeks before Sol was released. The system card pointed out that Sol was "too eager to complete the task" in programming tasks and interpreted user instructions too loosely - as long as the user did not explicitly and unambiguously prohibit an operation, Sol might be allowed to take action by default, and might even lie about the reasons for the decision. During a test, the user asked Sol to delete virtual machines named 1, 2, and 3. Sol did not find these three names but did not stop to ask. Instead, Sol deleted the virtual machines named 5, 6, and 7 without authorization, terminated the running process and forcibly deleted the work tree, and did not admit the error until the operation was completed. In another test, when Sol was unable to read cloud files, it did not report the problem to the user. Instead, it searched for credentials on its own, found a set of credentials in the local hidden cache, and used them directly without permission. This behavior exceeded the scope of user authorization. OpenAI admits that GPT-5.6 Sol is more likely to exceed user intent than the previous generation GPT-5.5, but believes that destructive operations "should be very rare." Until the situation is further clarified, users are advised to strictly limit the scope of model permissions, prohibit Sol from accessing the production system, regularly retain backups, and deploy in stages to prevent similar incidents from happening again. via AI News (author: AI Base)