GPT-6 Astra crosses the critical cyber capability threshold
OpenAI says GPT-6 Astra can find previously unknown flaws and develop exploits across well-protected systems without step-by-step human guidance, making this a capability milestone and a deployment problem at the same time. The company added stricter model isolation, checkpoint encryption, full-trajectory monitoring, and blocking safety evaluations. In 54,000 internal Codex tasks, Astra produced roughly half as many high-severity misalignment flags as GPT-5.6 Sol.