OpenAI locks down Astra over potential critical cyber capabilities

OpenAI locks down Astra over potential critical cyber capabilities OpenAI’s internal evaluation of its upcoming model, Astra, found significant advances in agentic coding and cybersecurity, leading the company to conclude that it cannot rule out the model reaching the critical capability level for cybersecurity under its Preparedness Framework. The Preparedness Framework, first published in December 2023, outlines how OpenAI evaluates frontier AI risks and determines the safeguards required before deploying increasingly capable models. It identifies high-risk capability areas, including cybersecurity, biological and chemical threats, harmful persuasion, and AI self-improvement, and requires models to undergo safety evaluations before deployment. If a model is assessed as posing an unacceptable level of risk, OpenAI says it will delay deployment or introduce additional safeguards until the risk is reduced. How OpenAI assesses critical cyber capabilities Under the Preparedness Framework, OpenAI classifies a model as having critical cybersecurity capabilities if it can independently discover previously unknown software vulnerabilities in secure systems or plan and execute sophisticated cyberattacks against well-protected targets with little or no human guidance. The company is following the same approach it adopted in 2025 when its AI models began demonstrating advanced biology capabilities. “While we continue to benchmark and assess this model, our preliminary evaluations indicate strong enough performance that we cannot rule out the critical capability level at this time. Astra is an upcoming model and was not involved in exploiting Hugging Face,” OpenAI said. The company said that this is a preliminary assessment and that Astra has not been classified as a critical cybersecurity model, and it is continuing evaluations before making a final determination under its Preparedness Framework. Enhanced security controls OpenAI has strengthened safeguards and security controls to support the deployment of models with these capabilities. During Astra’s development, it began implementing stricter security measures for higher-capability models and related activities, including isolated testing environments, restricted network and tool access, enhanced protection and encryption of model weights, additional monitoring and detection systems, and sandboxed execution. The company said it has paused activities involving Astra that do not meet these enhanced security requirements and introduced comprehensive monitoring to detect risky actions and signs of misalignment across agentic applications. Before deploying Astra, OpenAI will ask government agencies and independent AI safety organizations to evaluate the model’s cybersecurity capabilities. It will also provide external testing partners with guidance and security measures to enable the safe evaluation of the model’s higher-risk capabilities. Download: The high-performance team playbook