OpenAI's GPT-6 Astra is the first model to hit its own 'Critical' cyber threshold

OpenAI has given general availability to GPT-6 Astra, which the company calls its most capable broadly deployed model, and says it is the first model to cross the Critical cybersecurity capability threshold under OpenAI's own Preparedness Framework. That classification is OpenAI's internal risk grading, not a rating from an independent regulator, and it changed how the company shipped the model.

What "Critical" means here

OpenAI's framework defines Critical cyber capability as a model that, given the right tools and access, can find previously unknown security flaws and develop new ways to exploit them across well-protected systems largely without a person guiding each step. Astra is the first OpenAI model to cross that line; the prior model, GPT-5.6 Sol, was classified only as High. OpenAI says it delayed Astra's release specifically to add safety testing once it determined the model had reached this level, and shipped the commercial version with restrictions on its most advanced offensive capabilities. A less-restricted variant is available only through OpenAI's Daybreak program, aimed at cybersecurity defenders doing vulnerability validation, malware analysis and detection engineering.

Vendor-stated safety claims

OpenAI also states Astra is more robust against jailbreak attempts than GPT-5.6 Sol, including over longer conversation trajectories, and reports roughly half as many higher-severity misalignment flags across an internal simulation of more than 54,000 Codex tasks. These are OpenAI's own internal test results, not independently audited figures, and should be treated as vendor-reported context rather than a substitute for a buyer's own evaluation.

What MSP and enterprise buyers should check

Before enabling Astra-class access for customer environments, the useful questions are practical rather than rhetorical: what access gating applies, whether agent actions and reasoning trajectories are logged, what the refusal and escalation path looks like for flagged high-risk users, and whether "Critical cyber capability" changes anything about acceptable use in a managed environment. None of that marketing language is permission to run offensive security workflows outside an approved, defensive scope.

Sources: OpenAI safety overview, OpenAI announcement, NBC News.