OpenAI has slowed down development of its next-generation AI model, Astra, due to internal evaluations revealing advancements in agentic coding and cybersecurity capabilities that could push it into a "Critical" risk zone This article explores openai preliminary testing. . The company reviewed results from recent internal testing alongside external expert assessments, concluding they cannot currently rule out critical cyber capabilities based on their Preparedness Framework, which OpenAI has used since December 2023 to track and respond to AI advancements in biology, chemistry, cybersecurity, and self-improvement.
Earlier models, including GPT-5.6-Sol, were evaluated for frontier cyber capabilities and rated at the "High" threshold rather than "Critical." OpenAI's preliminary testing suggests Astra’s performance is strong enough that this threshold cannot be excluded, prompting the company to disclose the finding publicly in the interest of transparency with the safety and security research community. It has introduced stricter security measures for high-capability models, including isolated testing environments, restricted network access, stronger weight protections, enhanced monitoring systems, and sandboxed execution.
A universal monitoring system is now in place across all uses of the model, covering both training and evaluation phases.
