According to CoinMeta, OpenAI has officially confirmed that the upcoming Astra has met its “critical” network security capability threshold, becoming the first model to be rated at this level. Astra is capable of autonomously identifying unknown vulnerabilities and developing attack methods without human guidance. In exploitbench, Astra received a 100% rating. OpenAI also conducted internal tests on 20 recently disclosed high-risk vulnerabilities and discovered and exploited two zero-day vulnerabilities during these tests. Expert tests have shown that Astra can break through browser sandboxes and execute commands on the host machine; it can even attack fortified operating systems, elevating the privileges of ordinary users to root. The advanced version of Daybreak Blue will be initially available to a limited number of testers at the time of its official launch, before being rolled out to defense teams. The Astra training program, which was suspended due to a Hugging Face security incident, has also resumed, and OpenAI restarted large-scale cutting-edge RL training on August 28th.