OpenAI Pauses Astra Over Critical Cybersecurity Risks

OpenAI has hit pause on parts of development for Astra, its next AI model, after internal tests turned up cybersecurity capabilities the company says it cannot fully account for. The decision, made "last night" according to OpenAI, marks the first time the company has flagged one of its own systems as potentially reaching "Critical," the highest tier in its Preparedness Framework. Every prior model, including GPT-5.6-Sol, topped out at "High."

The framework exists to catch AI systems capable of acting on their own in ways that could cause serious harm, and cybersecurity is one of the categories OpenAI tracks most closely. A "Critical" rating there would mean a model can independently plan and carry out cyberattacks without a human directing each step. OpenAI introduced Astra only last week, and reports had pointed to a possible release as early as next month.

Now the company says it's rolling out tighter security controls, isolated testing environments, and automated systems designed to shut down risky behavior the moment it's detected. The trigger for that response traces back to something OpenAI's own internal testers ran into first.