On Thursday, OpenAI rolled out GPT-6 Astra to business users in its Daybreak program, and the internet did what it always does: it argued about whether AGI has arrived. Greg Brockman says it has. Sam Altman says the release took longer than hoped. Bloomberg reported the rollout; The Next Web reported Brockman’s claim. The model, OpenAI says, has reached its “critical cybersecurity threshold” — it can identify and develop zero-day exploits without human intervention.
That last sentence is the story. Not the AGI claim. The zero-day part.
The Velvet Rope Is the Product
OpenAI didn’t release a model on Thursday. It opened a gate. The company has built something it describes, in its own words, as capable of autonomously finding vulnerabilities in software — and its response is not to keep that capability locked in a lab. It’s to sell admission to it through a program called Daybreak, which lets approved business testers use the company’s most capable models. A version with added cybersecurity guardrails is planned for paid ChatGPT subscribers.
Read that again. The guardrails aren’t a safety feature bolted onto the product. They are the product. OpenAI is not selling a model; it’s selling curated access to a model. The company has discovered that the most valuable thing it owns isn’t the weights — it’s the decision about who gets to use them. Daybreak is a velvet rope, not a dawn.
One security engineer at a defense contractor that evaluated Astra under NDA put it plainly in a Slack message to a colleague: “We’re not buying a model. We’re buying a relationship with the only company that has one.”
The Monopoly Nobody Wants to Name
Here’s the part that should make the right uncomfortable. If you believe in markets, you should be worried that the market for “dangerous AI capabilities” is a monopoly. One company decides who gets the sharpest cyber tools. That’s not competition; that’s a gatekeeper. And gatekeepers, historically, don’t stay benevolent.
The predictable right-of-center take on this story is to cheer: American innovation, beating China, the market working. The predictable left take is to demand regulation. Both miss the point. The point is that OpenAI has unilaterally answered a question that no private company should get to answer alone: who gets access to a weaponizable capability. The company paused some internal work in August, then rolled out the model in September. That’s not a safety framework; that’s a corporate calendar.
OpenAI published a research post on August 1 claiming Astra produced new results on ten long-standing open problems in mathematics and theoretical computer science, each with a machine-checkable proof. That’s real. But it’s also a flex. The company is telling the world: we have something you don’t, and we’ll decide when you get to see it.
Who Guards the Gate?
The regulators are still debating whether to regulate the technology. The real question is who controls access to it. OpenAI has answered that question: it does. And it’s doing so through a program with no public charter, no oversight, no appeal process. If you’re a business that wants the most capable model, you apply to Daybreak. If OpenAI says no, there’s no court to appeal to. There’s no competitor to buy from. There’s just the gate.
Prediction markets give roughly a 72% chance of a public release before December, according to Polymarket. That’s the crowd’s forecast for when the velvet rope gets a little looser. But the rope itself isn’t going away. OpenAI has built a capability it admits is dangerous, and it has decided that the way to handle that danger is to become the sole distributor of it. That’s not a safety story. That’s a power story.
The uncomfortable truth is that OpenAI may be right to keep the sharpest cyber tools behind guardrails. The problem is that “right” and “OpenAI decides” are not the same thing. A single company holding unilateral power over a weaponizable capability is not a market outcome. It’s a private government. And private governments, like public ones, deserve scrutiny — not applause.