OpenAI's Astra: AI Model Hits Critical Cyber Limit

OpenAI is slowing down work on its upcoming AI model, Astra. This is because internal testing suggests it has reached a cybersecurity limit no previous OpenAI model has hit. The company states it "cannot rule out critical cyber capabilities" in Astra. They have paused some internal activities and are expanding testing with tighter controls. Here's the thing: under OpenAI's Preparedness Framework, a model is "Critical" if it can identify and develop functional zero-day activities in many real-world systems without human involvement. It can also be Critical if it performs a novel end-to-end attack against a hardened target from just a high-level goal. Preliminary results were strong enough that OpenAI couldn't confidently place Astra below this level. What's interesting is that previous models were measured at the "High" cybersecurity level. Astra is now being tested in isolated environments with tighter network limits. The bottom line: for developers, the concern is that the skills making coding agents useful could be turned against the software they were built to work on. This highlights the growing security challenges with advanced AI.