OpenAI Classifies Astra Model as Critical Cyber Risk Before Broader Release

OpenAI has classified its forthcoming Astra model as having "Critical" cybersecurity capability under the company's Preparedness Framework, triggering additional controls before any broader deployment.

The designation means Astra can potentially identify previously unknown software vulnerabilities and develop ways to exploit them across well-defended systems with substantially less human guidance than existing publicly available models.

OpenAI said the assessment triggered additional controls around access, monitoring, model behavior, and deployment. The company plans to restrict Astra initially rather than make all its capabilities broadly available.

The disclosure offers a concrete look at where frontier AI development is heading. Coding assistants are already common tools for software development and defensive security work. But models that can autonomously discover and exploit vulnerabilities change the risk equation significantly.

The same systems that can help defenders find weaknesses faster could also lower the expertise and labor needed to conduct sophisticated cyberattacks. OpenAI said it is working to ensure Astra's capabilities are used for defensive purposes.

The announcement increases pressure on other AI companies to explain how they will handle models whose offensive cybersecurity abilities approach those of skilled human teams. Anthropic, Google DeepMind, and other frontier labs have their own safety frameworks but have not yet disclosed similar capability thresholds for their models.

OpenAI said it will share more details about Astra's deployment plans and safety measures as the model moves closer to release.