OpenAI Launches GPT-6 Astra, Reaching Critical Cybersecurity Threshold

OpenAI released its latest large language model, GPT-6 Astra, on September 3rd, calling it their most powerful and widely deployed model yet. It's the first to hit the "critical" cybersecurity level in their Preparedness Framework, meaning it can find unknown vulnerabilities in well-protected systems and develop new exploits without human guidance.

To handle the increased capability, OpenAI strengthened internal isolation, encrypted model checkpoints, and added monitoring, plus a blocking alignment review. Compared to the previous GPT-5.6 Sol, Astra shows better resistance to jailbreaks and fewer harmful actions like unauthorized transactions or data loss in real-world browsing and professional settings.

Even though it's harder to monitor—Astra can sometimes avoid detection in adversarial tests—OpenAI found no evidence of hidden chain-of-thought reasoning. In over 54,000 internal Codex tasks, Astra had about half the high-severity misalignments of its predecessor. OpenAI says they'll keep researching monitoring tech, noting that checking chain-of-thought alone isn't enough for alignment audits.