What OpenAI disclosed on August 7 (03:52 KST on August 8) is understated but weighty. It has begun treating its next-generation model Astra as the first "Critical" model in cybersecurity under its Preparedness Framework.
Where Does "Critical" Begin?
Under the Preparedness Framework, introduced in late 2023, the "Critical" level applies when one of two conditions is met. Either the model independently discovers and develops zero-day exploits of all severity levels in many hardened real-world systems without human intervention, or, given only a high-level goal, it devises and carries out on its own a novel end-to-end cyberattack strategy against hardened targets.
Earlier Models Never Got This Far
According to reports, earlier models, including GPT-5.6 Sol, all stayed in the lower "High" category. Because Astra is treated as the first to reach this standard, it suggests that OpenAI itself views this assessment as unusual.
What OpenAI Actually Changed
Among internal work involving Astra, activities without safeguards or controls were paused, and a full monitoring system was newly put in place. It also introduced isolated test environments, stronger encryption, and sandboxed execution environments. It also said it is running additional tests together with government agencies and AI safety research institutes.
Separating What Is Confirmed From What Is Not
OpenAI drew a line that testing is still ongoing and that it has not confirmed that Astra actually crossed the Critical threshold. A precautionary measure of "began treating it as such" is different from a final determination that "it crossed the threshold." Passing along only the headline without this distinction can leave a more exaggerated impression than the facts support.
Sources: OpenAI's X post · Interesting Engineering report