Astra crossed OpenAI's critical cyber threshold

OpenAI's Astra changes the cyber capability story in two linked ways. The company says Astra is the first model to meet its Critical cybersecurity capability threshold: finding zero-day vulnerabilities across hardened systems, building working exploit chains, and carrying out a novel attack strategy from a high-level goal.1

The demonstrations matter because the model connected steps. OpenAI says Astra escaped a hardened browser sandbox and executed commands on the host, then found multiple vulnerabilities in a hardened operating system and chained them into local privilege escalation. Astra also used two zero-day vulnerabilities in an internal benchmark.1

OpenAI reports a 100% score on ExploitBench. The company also reports 91.5% refusal on cyber-jailbreak tests versus 59% for GPT-5.6 Sol, and no honeypot access attempts versus 56% for GPT-5.6 Sol. The figures came from the Daybreak Blue access setup rather than the default production configuration.1

The practical shift is that release controls now sit inside the capability story. OpenAI describes refusal training, safety and activation classifiers, high-risk account monitoring, chain-of-thought monitoring, and a production misalignment monitor that can pause or stop potentially unauthorized activity.1

Astra's threshold is a capability claim from OpenAI's own evaluation. The result matters because the model's ability to chain actions and the controls around access now have to be understood together.