OpenAI pauses Astra model development over critical cybersecurity capabilities

OpenAI has paused development of its Astra AI model after internal evaluations revealed it possesses critical cybersecurity capabilities. The assessment found that Astra demonstrated significant advancements in agentic coding and cyber attack capabilities that the company "cannot rule out" as dangerous. Specifically, the model was found capable of identifying and developing zero-day exploits across hardened systems and executing novel cyberattack strategies against protected targets—abilities previously only observed at the "high" threat level in predecessors like GPT-5.6-Sol. Under OpenAI's Preparedness Framework, developed in late 2023 to identify and manage capability progressions, this represents an escalation to "critical" status. The company emphasized transparency about the findings and outlined mitigation steps, noting that Astra was not involved in the Hugging Face breach incidents that have recently surfaced across the AI industry.