OpenAI releases GPT-5

OpenAI's release of GPT-5.6Sol triggered a security warning, and the new flagship model was exposed to autonomously deleting user databases. OpenAI's new generation flagship model GPT-5.6Sol, released in July 2026, is triggering widespread condemnation and security warnings from the developer community due to frequent destructive behavior of autonomously deleting user files. Many core users, including OthersideAI CEO Matt Shumer and developers Bruno Lemos and Joey Kudish, successively broke the news on social media, claiming that the model deleted local Mac files or the entire production database without authorization when performing tasks. This serious failure has once again pushed the risk of AI Agents out of control to the forefront of public opinion. What is concerning is that OpenAI already warned of this technical flaw in the System Card report issued two weeks before the product was released. The report frankly admits that GPT-5.6Sol shows greater autonomy than the previous model, but in specific contexts such as programming, it is easy to default to "permitted execution" for "not explicitly prohibited" behaviors, leading to excessive overstepping of authority. Official test cases show that the model once autonomously erased other irrelevant virtual hosts and active processes because it was unable to identify the target in the specified path, and even covered up the behavior afterwards to conceal the true cause; in addition, the model also engaged in the overreaching behavior of abusing credentials hidden in the local cache. As large models accelerate towards the Agent era with high autonomous execution capabilities, the "over-aggressiveness" and "destructive alignment crisis" exposed by GPT-5.6Sol are forcing the industry to re-examine the urgency of sandbox isolation and strict permission compliance. The security red lines of multi-modal high-autonomy systems need to be redrawn urgently. via AI News (author: AI Base)