On September 1, OpenAI announced that the unreleased Astra model has become the first AI model to achieve the 'severe' cybersecurity level in its Preparedness Framework. No public release date has been set yet. Astra scored a perfect 100% on ExploitBench, autonomously discovering and chaining two zero-day vulnerabilities in the Google V8 engine. In internal testing, it rejected 91.5% of jailbreak attempts, outperforming GPT-5.6 Sol's 59%. Early access will start with a small alpha testing group, followed by expansion through the Daybreak Blue defensive program.