OpenAI on Friday said its upcoming artificial intelligence model, Astra, has demonstrated significant advances in agentic coding and cybersecurity, prompting the company to conclude that it cannot currently rule out the model reaching the “Critical” cybersecurity capability threshold under its Preparedness Framework. The company said its latest internal evaluations, conducted over the past few days along with assessments by experts, showed a notable improvement in the model’s ability to perform cybersecurity-related tasks. OpenAI said it was sharing the findings to maintain transparency with the public, governments and the broader AI safety and security community. Under OpenAI’s Preparedness Framework, a model reaches the Critical cybersecurity threshold if it can identify and develop functional zero-day exploits across severity levels in multiple hardened, real-world critical systems without human intervention. The threshold also includes the ability to develop and execute novel, end-to-end cyberattack strategies against hardened targets based only on a high-level objective.
OpenAI clarified that Astra is an upcoming model and was not involved in the exploitation of Hugging Face. The company added that its preliminary evaluations are still ongoing, but performance has been sufficiently strong that Critical-level capability cannot be ruled out at this stage. In response, OpenAI said it has expanded robustness testing of Astra’s safeguards and security controls to ensure they are suitable for models with potentially advanced cyber capabilities. The company is implementing stricter security measures for higher-capability models, including isolated testing environments, restricted network and tool access, stronger protection and encryption of model weights, enhanced monitoring and detection systems, and sandboxed execution. OpenAI said it has also paused internal Astra-related activities that do not yet meet the strengthened security requirements. Universal monitoring for risky actions and potential misalignment has been implemented across Astra’s agentic applications, including training and evaluation. These monitoring systems assess the model’s chain of thought and can trigger a security response to review or interrupt high-risk activity. The company also plans to work with relevant government agencies and selected AI safety organisations to test Astra’s capabilities, while providing recommended security controls to third-party testing partners conducting higher-risk evaluations. OpenAI said its Preparedness Framework had previously guided its response to emerging capabilities in areas such as biology. The company said it wants advanced cyber-capable models to help defenders identify and address vulnerabilities before malicious actors can exploit them, while stressing the need for responsible deployment of increasingly capable AI systems.
