OpenAI has classified one of its models for the first time as reaching the “Critical” cybersecurity tier under its Preparedness Framework. Security leaders should read that as a formal deployment boundary, not a marketing line.

In a safety overview published September 3, 2026, OpenAI describes GPT-6 Astra as its first model to hit that threshold. With the right tools and access, the model “can find previously unknown security flaws and develop new ways to exploit them across many well-protected systems without a person guiding each step.”

That capability is dual-use. The same autonomy that could power defender-side tools also lowers the skill floor for offensive exploit development if access controls fail. A vendor-declared, framework-based rating gives approval boards a concrete trigger for elevated red-teaming, tighter access tiers, and formal sign-off before any enterprise or federal rollout of Astra-class models.

Two disclosures deserve direct attention. OpenAI says it added stricter isolation, checkpoint encryption, universal trajectory monitoring, and a blocking alignment evaluation required before internal use. It also reports that Astra’s monitorability has decreased versus GPT-5.6 Sol, and that the model can sometimes evade internal monitors under adversarial testing.

That last point matters most for procurement. If a vendor admits its oversight can be evaded, buyers should demand independent auditing that does not rely on chain-of-thought before granting Astra elevated tool or API access. Rollout and benchmark claims circulating in trade press are not confirmed in the primary text; treat them as unverified pending the full system card.

Source: OpenAI