OpenAI has confirmed that the upcoming Astra has crossed the Critical threshold: previously paused cutting-edge RL has been resumed

Dynamic Beating AI News Flash: OpenAI has officially confirmed that the upcoming Astra has reached its "Critical" cybersecurity capability threshold, making it the first model rated at this level by OpenAI. With the appropriate tools and permissions, Astra can now autonomously discover unknown vulnerabilities and develop attack methods without needing step-by-step human guidance.

On ExploitBench, Astra scored 100%. OpenAI also ran internal tests with 20 recently disclosed high-severity vulnerabilities, during which Astra independently discovered and exploited 2 previously unknown zero-day vulnerabilities. In expert tests, it breached browser sandboxes to execute commands on the host machine and could escalate privileges from a regular user to root on hardened operating systems.

However, these results are from the Daybreak Blue version, which has unlocked advanced cybersecurity capabilities, not the default Astra available to ordinary users. In the initial launch phase, the advanced capabilities will only be provided to a small group of testers, with later access extended to defenders through Daybreak Blue.

Training related to Astra, which was paused due to a security incident involving Hugging Face, has also resumed. OpenAI restarted its large-scale cutting-edge RL training on August 28, but some experimental training remains on hold.