OpenAI has released GPT-6 Astra, an AI model that independently discovered two previously unknown zero-day vulnerabilities and demonstrated critical-level cyber capabilities—including arbitrary code execution, privilege escalation, and sandbox escape—across multiple test benchmarks. The model significantly outperformed its predecessor (GPT-5.6 Sol) in cyber exploitation tasks, achieving 100% on public exploit conversion tests and 99.2% on binary reverse-engineering with up to four attempts. Due to these findings, OpenAI has restricted advanced cyber features at launch and plans a phased rollout through its Daybreak enterprise program, prioritizing defensive use cases (code review, patching) while initially blocking offensive capabilities like proof-of-concept exploit generation. This represents the first OpenAI model classified as "Critical" under its risk assessment framework.