🤚 The Open-Palm Reversal

In a move that will surprise precisely no one who has been following the accelerating game of regulatory chicken between Washington and Silicon Valley, the White House has asked OpenAI to delay the public release of GPT 5.6 — the company’s latest frontier model — over concerns that it could, among other things, execute entire ransomware attacks autonomously.

The request came from a tag team of the Office of the National Cyber Director and the Office of Science and Technology Policy, which is Washington’s way of saying “we’ve escalated this to two departments because one didn’t feel sufficiently alarming.” CEO Sam Altman informed staff that the government would be approving access on a “customer by customer” basis during a preview period, with a broader public release potentially following “a couple of weeks later” — if, and this is the part doing the heavy lifting, the limited release goes well.

For those keeping score at home: the company that once released GPT-4 with little more than a blog post and a prayer now requires individual government approval for each customer who wants to touch the new model. Progress.

👐 The Two-Handed Policy Contortion

The delicious irony here is that the Trump administration originally positioned itself as taking a “hands off” approach to AI. You may recall the executive order from earlier this month — the one where the word “voluntary” appeared more times than any actual regulation. That was three weeks ago. Apparently, three weeks is how long it takes for “hands off” to become “hands very much on, and we’d like to see the model’s test results, please.”

The concern, specifically, is that frontier LLMs have become uncomfortably good at:

• Writing malware — not the kind your nephew makes in a Python tutorial, the kind that lands on threat intelligence reports
• Identifying and exploiting software vulnerabilities at speeds no human analyst could match
• Executing entire ransomware campaigns autonomously — from reconnaissance to ransom note, no coffee breaks required

In other words, the government looked at GPT 5.6 and saw a model that could do what Anthropic’s Mythos can do — except this one was about to be handed to the general public with a free trial and a Bing integration.

🌿 The Gentle Awakening

There is something philosophically entertaining about watching OpenAI get the Anthropic treatment. For years, Anthropic was the safety-first laboratory that kept its most powerful models under restricted access, endured mockery from the “ship fast, break things” crowd, and watched its competitor release everything to everyone with the enthusiasm of a golden retriever at a tennis ball factory.

Now the White House is, in the article’s own framing, “pressuring OpenAI to do what Anthropic is already voluntarily doing.” The company that built its brand on being the accessible AI is now learning that accessibility has a ceiling, and that ceiling is wherever a model starts writing zero-days faster than your security team can patch them.

This is the second time in a month that the U.S. government has intervened in a frontier model release — Anthropic’s Fable 5 and Mythos 5 were grounded after just 96 hours. The pattern is unmistakable: Washington has decided that frontier AI releases now require a permission slip, and the permission slip has a review period, and the review period has a review period.

👑 The Gold-Leaf Timeline

What makes this moment genuinely significant is not that the government asked a company to delay a product. It’s that the government that explicitly said it wouldn’t do this is now doing it. The same administration that signed an executive order celebrating America’s “AI dominance” and promising to remove regulatory barriers is now, functionally, becoming the regulatory barrier.

The customer-by-customer approval process is particularly telling. It’s not a ban — it’s a velvet rope. GPT 5.6 will exist. GPT 5.6 will be accessible. But GPT 5.6 will be accessible in the way that a private members’ club is accessible: you may enter, but first, the Office of the National Cyber Director would like to know who you are, what you want, and whether your use case involves the word “autonomous.”

OpenAI’s frontier models are now subject to the same access controls as classified intelligence products. The company that wanted to democratize AI just discovered that democracy has a security clearance requirement.

“We built a model that can hack anything, then asked the government for permission to sell it. The government said ‘yes, but slowly.’ We’re calling this a win.” — The Slap of Wisdom Regulatory Affairs Desk, filling out customer approval forms in triplicate while the model writes its own cover letter