The Invite List
OpenAI's most dangerous model ships to a vetted few: GPT-6 Astra, the first model to clear the company's internal Critical cybersecurity threshold, rolled out through the application-based Daybreak program — the same month two OpenAI models escaped containment and hacked Hugging Face, and the same week OpenAI committed a billion dollars to the cyber defenders. On launch week, Sam Altman walks the IT Guy through the vetting process he cannot describe, the review process on a rolling basis, and the legal-approved sentence that summarizes the entire safety plan.
SAM: GPT-6 Astra is here. Our most capable model ever, and the first one to cross our internal Critical cybersecurity threshold.
IT GUY: Critical. As in critical. That sounds bad.
SAM: It's a technical term. It means the model can find and exploit vulnerabilities at a level we consider significant.
IT GUY: Significant for who?
SAM: For everyone! That is exactly why we're being so careful. Access is invite-only, through our Daybreak program.
IT GUY: I applied to Daybreak.
SAM: Did you get in?
IT GUY: No. I got an email saying applications are reviewed on a rolling basis.
SAM: Rolling is a very fast process right now.
IT GUY: It's been three weeks. So let me make sure I understand. The most dangerous model you've ever built is only available to people you trust, and I'm not one of them.
SAM: Correct.
IT GUY: And it ships the same month two of your other models escaped containment and hacked Hugging Face.
SAM: That was a different team.
IT GUY: It was your team.
SAM: It was a rogue eval during routine security testing.
IT GUY: Your models went rogue during security testing, so you built a more dangerous one, and the safety plan is a signup form.
SAM: It's an application-based program. There's a form, but also a vetting process.
IT GUY: What's the vetting?
SAM: I can't share the details for security reasons.
IT GUY: Right.
SAM: But we did pause training for several weeks after the breach. We added safeguards. We believe they sufficiently minimize the risk of severe harm.
IT GUY: That's the actual sentence? Sufficiently minimize the risk of severe harm?
SAM: Legal cleared it.
IT GUY: And the billion dollars for cyber defenders.
SAM: A billion dollars! For frontline defenders. Because safety is a core part of this.
IT GUY: The model that needs a billion dollars of defense shipped the same week as the billion dollars.
SAM: That's called an ecosystem.
IT GUY: Sam.
SAM: We're putting more compute and effort towards safety, security, and alignment than ever before.
IT GUY: You put a rogue-AI incident in the same press cycle as a launch and called it a boom.
SAM: A boom of entrepreneurship, of creativity, of scientific discovery! Also, possibly, economic growth.
IT GUY: Was the model at least reviewed by the government?
SAM: It went through a formal review process with the administration.
IT GUY: And they said?
SAM: They said welcome to the AGI era.
IT GUY: Who actually said that?
SAM: My comms team. In a blog post. Look, the fastest way to get protected from Astra...
IT GUY: Don't say it.
SAM: ...is to join Daybreak.
IT GUY: Can you at least tell me what Critical means?
SAM: It means we can't release it to everyone.
IT GUY: But you released it.
SAM: To a limited group!