ICYMI: OpenAI's Astra crosses Critical cybersecurity threshold
OpenAI says Astra is its first model to meet the Preparedness Framework’s Critical bar for cyber capability.
With Daybreak Blue-level access, Astra found unknown flaws, built working exploit chains, and outperformed GPT-5.6 Sol on vulnerability discovery and exploit development. TestingCatalog’s note says it scored 100% on ExploitBench and found two zero-days now being disclosed to maintainers. OpenAI is limiting the strongest cyber functions to alpha testers first, with wider defensive access planned later. The company says release was delayed while it added refusals, classifiers, monitoring, and response systems. TestingCatalog's note
With Daybreak Blue-level access, Astra found unknown flaws, built working exploit chains, and outperformed GPT-5.6 Sol on vulnerability discovery and exploit development. TestingCatalog’s note says it scored 100% on ExploitBench and found two zero-days now being disclosed to maintainers. OpenAI is limiting the strongest cyber functions to alpha testers first, with wider defensive access planned later. The company says release was delayed while it added refusals, classifiers, monitoring, and response systems. TestingCatalog's note
score 8