Critical Acclaim
OpenAI’s next model is the first in company history to hit the Critical threshold for cybersecurity capability — meaning it is very good at breaking into things, and access will be handled the way normal companies handle donuts: by application. The persona below is invented. The rating is not.
The Award Show
Every industry handles dangerous products differently. Airlines ground the fleet. Pharma recalls the lot. The software industry throws a keynote, publishes a tier chart, and announces that the most hazardous model in company history is — and we want to be precise here — available by application.
It is worth sitting with the rating system itself. Low, Medium, High: retired, like yesterday’s pricing page. Critical: new, glowing, at the top. A capability ladder where every rung is a press release and the top rung is a warning label, and the warning label is somehow also the marketing.
Access, Apparently
The safeguard is an application. You write an essay about who you are and what you intend to do with frontier cyber capability, and it gets read — thoughtfully, presumably, the way terms of service are written. The model that cannot be ruled out as capable of executing complex cyberattacks with minimal human input is gated behind the same technology as a farmers market stall: trust, but verify, but make it a Google Form.
And the timing discipline is real: announced on a Wednesday. Because Friday announcements are for breaches. Wednesday is for things that haven’t breached anything yet. That is the bar. That is the framework.
What We Cannot Confirm
- That anyone at OpenAI described the Critical rating as “a franchise record.” (Invented.)
- That access applications are graded on a rubric where honesty is worth extra credit. (Invented, and honestly, plausible-adjacent.)
- That the Preparedness Framework chart doubles as the launch slide. (Invented. The chart is real.)