OpenAI confirms its upcoming Astra has crossed the critical threshold: the previously paused frontier reinforcement learning (RL) work has been restarted.

Beating AI Express: OpenAI has officially confirmed that its upcoming model Astra has reached the "Critical" cybersecurity capability threshold, becoming the first model rated at this level by OpenAI. When provided with appropriate tools and permissions, Astra can independently identify unknown vulnerabilities and develop attack methods without step-by-step human guidance, scoring 100% on ExploitBench. OpenAI conducted internal tests using 20 recently disclosed high-severity vulnerabilities, during which Astra also independently discovered and exploited two previously unknown zero-day vulnerabilities. Expert tests showed Astra can break through browser sandboxes to execute commands on host machines, and attack hardened operating systems to escalate privileges from regular users to root. These results are from the Daybreak Blue version, which has advanced cybersecurity capabilities, not the default Astra version for regular users. During its initial launch phase, advanced capabilities will only be available to a small number of testers, and will later be opened to defenders via Daybreak Blue. Astra-related training, suspended due to a Hugging Face security incident, has resumed: OpenAI restarted the large-scale frontier reinforcement learning (RL) training suspended on August 28, though some experimental training remains suspended.