OpenAI Delays Astra Release: Cyber Risks & Gov't Testing

OpenAI has slowed the release of its new model, Astra, after internal tests showed it has cybersecurity capabilities that could pose digital security risks. The company confirmed this decision, stating that Astra may reach the "Critical" threat level in its Preparedness Framework. A Critical designation means an AI model could create functional zero-day exploits or carry out complex cyberattacks from broad instructions. Previous models, like GPT-5.6-Sol, were rated "High" risk. OpenAI cannot rule out Astra having "Critical" capabilities, prompting these precautionary measures. The company has suspended internal development projects for Astra that don't meet new security standards. They are implementing stricter operational controls, including enhanced encryption for model access and testing in secure, sandboxed environments. OpenAI will also work with government authorities and AI safety groups for external evaluations. This matters because it highlights the serious security considerations in developing advanced AI.