OpenAI Hits Cybersecurity Red Line, Partially Halts Rollout of Next-Gen Astra Model

On August 10, OpenAI confirmed that it has paused parts of its work on Astra, the company’s next-generation large language model, after internal safety testing raised red flags about its cyber capabilities. According to Axios, OpenAI’s “Preparedness” framework classified some of Astra’s potential offensive security skills as approaching a critical risk threshold, prompting the company to slow deployment and tighten internal access while additional evaluations are carried out. At the same time, OpenAI introduced GPT‑5.6‑Cyber, a new model explicitly optimized for cybersecurity defense. During testing it handled the vast majority of advanced security tasks, but—unlike Astra—was judged to stay below the company’s highest concern tier for offensive use, allowing OpenAI to release it with fewer restrictions. The move is one of the clearest examples so far of a frontier AI lab voluntarily trading speed for safety, and it puts pressure on rivals and regulators to define what level of cyber capability should trigger mandatory safeguards or oversight. Whether Astra ultimately ships in a more constrained form, or becomes a template for stricter industry norms on dual‑use AI, will be a key signal for how seriously the sector takes its own risk frameworks in practice.