Frontier AI crosses the “Critical” cyber line

OpenAI’s GPT-6 Astra system card, published September 3, says the model “is a significant step up in cyber capabilities and meets our Critical threshold.” It is the first OpenAI model to reach that threshold. OpenAI explains what that means: “With the right tools and access, GPT-6 Astra can find previously unknown security flaws and develop new ways to exploit them across many well-protected systems without a person guiding each step.” The model nonetheless rolled out broadly to paid ChatGPT plans, the API, and Amazon Bedrock. It costs $10 per million input tokens and $50 per million output.

On September 22 Anthropic released Claude Opus 5.5, which it says “performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5,” at $4 and $20 per million tokens. Because the model “has extremely strong cyber capabilities,” Anthropic applied safeguards under which routine bug-fixing still works, “but most cybersecurity tasks will be re-routed to Opus 4.8.” Vetted defenders get a separate path through an expanded Cyber Verification Program.

Two risks, one month

The first risk is on the threat side. Autonomous discovery and exploitation of unknown flaws is now a commercial capability, with the labs’ safeguards as the main brake on misuse. For a community bank, the exposed surface is mostly someone else’s code, such as the core, the online-banking platform and the payment processor. That makes a provider’s patch cadence and incident transparency a first-order control.

The second is on the vendor side. Safeguard routing means the model a vendor contracts for is not always the model that answers. By design, some requests go to a different, older model. Enterprise access is also a configuration decision, because ChatGPT Enterprise administrators must switch Astra on for their organizations. A vendor attestation that names one model and one version no longer describes what actually runs.

When the model can find the zero-day, the question for your vendors is no longer whether they use AI. The question is how fast they patch, and how fast they tell you.