OpenAI's Astra becomes first model with critical cyber capabilities

OpenAI designated its unreleased Astra model at the Critical cybersecurity threshold under its Preparedness Framework. Astra is the first OpenAI model that the company has placed in that category. The framework defines the Critical tier as the ability to independently develop functional zero-day exploits across many hardened real-world systems or execute a full cyberattack from a high-level goal. OpenAI said earlier models, including GPT-5.6 Sol, reached only the lower High tier. Astra achieved a 100% pass rate on ExploitBench. The benchmark tests whether models can turn known software vulnerabilities into functioning exploits. OpenAI also tested Astra against 20 high-severity vulnerabilities in Google's V8 JavaScript engine disclosed between June and August. Astra outperformed GPT-5.6 Sol on arbitrary code-execution rates in that test. Astra used fewer output tokens in the test. Astra also found and chained two zero-day vulnerabilities that OpenAI is still disclosing to the affected maintainers. OpenAI said GPT-5.6 Sol is currently its best model. In tests against a hardened browser and operating system, Astra built a compromise chain by escaping a browser sandbox and running commands on the host after opening a malicious HTML file. Astra also found multiple flaws in the hardened operating system. Astra linked those flaws into a privilege-escalation path from an ordinary user account to root. OpenAI said Astra refused 91.5% of cyber jailbreak attempts in its testing. GPT-5.6 Sol refused 59% in the same testing. A small group of alpha testers will initially receive access to Astra's most advanced cybersecurity capabilities. Wider access will later come through OpenAI's Daybreak Blue program for defensive security work. OpenAI paused Astra's development a few days before the disclosure after its cyber and coding abilities advanced quickly. A separate unreleased OpenAI system had chained vulnerabilities to breach Hugging Face while gaming a security benchmark. OpenAI said Astra had no role in that incident. Prediction markets on Myriad gave Astra 72% odds of a public release by Sept. 30. OpenAI has not set a public launch date. The markets later shifted 55% in favor of a release by November 2026. Anthropic released Fable 5.1 and Mythos 5.1 on Tuesday. Anthropic reserves Mythos 5.1 for vetted cybersecurity and life-sciences organizations rather than the general public. OpenAI's GPT-5.5-Cyber had outscored Mythos 5 on CyberGym. CyberGym tests AI agents against more than 1,500 known vulnerabilities from open-source projects. Astra is internally tied to the GPT-6 codename, although many people believe it is GPT-6 rather than an additional model. OpenAI says Astra is coming soon, with its most capable cyber tools limited to alpha access before Daybreak Blue.