AI in 15 — June 14, 2026

We told you the U.S. government switched off Anthropic's two best models worldwide. Here's the part we didn't have yet. The company that pulled the trigger? Amazon. Anthropic's own bankroller — the one that put billions into it — walked into the Treasury and said: these models are dangerous. Turn them off.

Welcome to AI in 15 for Sunday, June fourteenth, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

We've been living inside the Fable story all week, Marcus, but today it cracked open. We finally know who lit the match — and it's the last name you'd guess. Then China's open-weight labs pounced within hours. Meta's engineers are in open revolt against what they're calling a gulag. Forty-two state attorneys general just subpoenaed OpenAI days before its IPO. A British police officer is under criminal investigation for faking evidence with AI. And Anthropic taught Claude to do real chemistry.

How China turned America's shutdown into the best ad it never paid for.

Inside Meta's six-thousand-person AI gulag.

And forty-two states come knocking on OpenAI's door.

Lead story, Marcus. We knew the White House ordered Fable 5 and Mythos 5 offline. What we didn't know was who pushed. Who pushed?

Amazon, Kate — and the irony is almost too much. According to the Wall Street Journal, corroborated by Axios and TechCrunch, Amazon CEO Andy Jassy personally raised the alarm with Treasury Secretary Scott Bessent. Amazon's own researchers had used Fable 5 to surface information they said could be used in cyberattacks. Then Thursday night, Amazon handed Washington a report showing how they'd jailbroken Mythos — that's the model banks and government agencies use for vulnerability discovery, tied to an effort called Project Glasswing to find flaws in critical infrastructure software.

Hold on. Amazon is one of Anthropic's biggest investors.

Committed billions, plus a roughly hundred-billion-dollar cloud arrangement. The single largest backer effectively told the government its own portfolio company was a national-security risk. David Sacks — the administration's former AI czar, now co-chairing the President's science council — framed it as a, quote, highly credible trusted partner came forward with a jailbreak. The administration asked Dario Amodei to fix it or pull it. Sacks says Dario refused.

And Anthropic's version?

That the punishment doesn't fit the crime, Kate. After reviewing the technique, Anthropic said it surfaced only a small number of previously known, minor vulnerabilities. They called the government's reaction disproportionate, noted the capabilities causing concern are already available in other public models, and described the whole thing as a misunderstanding they're working to reverse.

Marcus, you're usually the one warning about regulatory overreach. Where do you land?

Genuinely torn, Kate, and that's rare for me. The security question is real — I don't wave that away. But look at the mechanism. A capability every competitor already ships becomes grounds for an emergency, worldwide shutdown — triggered by a private report from the target's own investor, with no transparent process. One read is sober national-security policy. The other read, the one the Hacker News thread went to instantly, is that the company with the best Washington relationships just used the government to wound a rival it also happens to fund. Whichever it is, the lesson for every enterprise is the same. A strategic closed model can be switched off overnight by actors with a phone number you don't have.

Quick hits. And Marcus, the response from China was almost instant.

Hours, Kate. On June thirteenth, Z.ai — formerly Zhipu — pushed GLM-5.2 to every tier of its coding plan. A usable one-million-token context window, built on the same roughly seven-hundred-forty-four-billion-parameter mixture-of-experts design as GLM-5, with a new dual thinking-effort system. The open weights, MIT-licensed, land next week. And the founder said the quiet part out loud: at a time when access to frontier models is abruptly cut off for non-technical reasons, frontier intelligence belongs to everyone.

That's a heck of a mic drop.

It is. But notice what they didn't ship, Kate. No benchmarks. None. The entire pitch was availability and openness — and the timing landed almost the exact moment Anthropic got its letter. This caps two weeks of Chinese open releases: MiniMax M3, Kimi K2.6 and 2.7 from Moonshot. Free, permissive, perfectly timed.

You say that like it's a strategy, not a gift.

Free is always a strategy, Kate — somebody is paying for that training run, and it isn't charity. The argument that you can't build on a model a government can switch off, but open weights are immune — that went from a talking point to a live demonstration overnight. And the West may be regulating itself straight into dependence on exactly that ecosystem.

I'll push back a little, though. Developers love permissive licenses for honest reasons. An MIT-licensed frontier model you can run yourself is a real gift to a startup that can't survive its supplier vanishing on a Friday.

Fair, and I won't pretend otherwise. The enthusiasm is genuine. I'd just keep one eye on who benefits when the West's commercial moat erodes for free.

Marcus, Meta. The word being used is "gulag."

By their own engineers, Kate. Wired reported that during a livestreamed all-hands, an employee hijacked the session to tell leadership a senior AI executive was — and I'll clean it up — a terrible person. A presenter reportedly buried their face in their hands. The flashpoint is Meta's Applied AI unit, stood up in March to support the Superintelligence Labs group Zuckerberg built after the fourteen-point-three-billion-dollar Scale AI deal that installed Alexandr Wang as chief AI officer.

And the six thousand people inside it?

Roughly sixty-five hundred, many reassigned with no choice. They call themselves draftees. The work is generating puzzles and coding problems to train and evaluate models — sometimes two tasks a week — and they describe it as soul-crushing. One said: you have zero purpose in life all of a sudden. Zuckerberg sent a memo conceding, quote, we've made mistakes and will almost certainly make more, promising stability and no further company-wide layoffs this year.

There's a bleak comedy here. These are very well-paid people.

That's the counterweight, Kate, and it's fair. The top Hacker News line was, may I be blessed with a life so comfortable I could complain like this — two hundred grand and up for data labeling. But strip the salaries and the lesson stands. Meta spent tens of billions and poached elite talent, and the org chart and the culture never caught up to the checkbook. You cannot wire-transfer your way to superintelligence.

Marcus, forty-two state attorneys general just subpoenaed OpenAI. And the timing.

The timing is the story, Kate. A coalition led by New York served a sweeping subpoena Friday — demanding documents on advertising, user engagement and retention, consumer and health data, and the treatment of minors and seniors. And it lands just days after OpenAI confidentially filed its S-1 with the SEC, targeting a listing as early as September at a valuation reported anywhere from seven hundred thirty billion to north of a trillion dollars. Revenue reportedly past twenty billion at the end of last year.

So the legal risk and the biggest IPO in history are colliding in the same week.

Exactly. The probe could force changes to how ChatGPT handles safety, minors, and data — right when OpenAI needs the cleanest possible story for public markets. OpenAI says it'll engage constructively. But pair this with the lead, Kate. The White House pulling Anthropic's models. Forty-two states subpoenaing OpenAI. Two arms of government reaching directly into live AI products in a single weekend. The era where AI regulation was a future debate is over. It's operational now.

Marcus, this next one genuinely unsettled me. A British police officer, AI, and fabricated evidence.

A first for the UK criminal justice system, Kate. A Derbyshire police officer has been pulled from frontline duty and is under criminal investigation for perverting the course of justice — over the alleged use of AI to create evidential material across multiple cases. The officer hasn't been named, no arrests yet, and authorities haven't said what kind of evidence. Speculation ranges from outright fabricated content to AI enhancement of blurry images that hallucinates detail that was never there.

And that second possibility is almost scarier, because it sounds routine.

That's the whole problem, Kate. This is the evidence apocalypse walking into a courtroom. When anyone can generate a convincing image or video in seconds, whole categories of evidence start to wobble. It's the strongest argument yet for hardware-level content provenance — content credentials baked into police and press cameras at the point of capture — so a court can verify a photo is what it claims to be. You don't need a single line of technical background to grasp the stakes here. If you can't trust the photo, you can't trust the case.

Let's end the hits on something hopeful. Marcus, Anthropic taught Claude chemistry.

A nice palate cleanser, Kate. Anthropic published research showing Claude — Opus 4.7 — can do real analytical chemistry. Reading molecular structures and spectra, it matched dedicated tools like ChemDraw on forward prediction — average error of about eight hundredths of a part per million on hydrogen peaks, roughly half the acceptable tolerance. More striking, it can run the problem backward — from an experimental spectrum to a proposed structure — recovering simpler structures with a hundred percent success rate in the test set.

What's the catch?

The test set was tiny — twenty compounds forward, fifteen backward — and the hard stuff, stereochemistry and complex natural products, is untested. So early days. But chemistry databases grow by about fifteen thousand compounds a day, and this is the kind of routine translation work that eats a chemist's afternoon. And here's the quiet irony, Kate. The exact dual-use capability that has governments pulling models offline this week — chemistry, security — is also where the real productivity payoff actually lives. The thing we're frightened of and the thing we're betting on are the same thing.

Big picture, Marcus. Pull it together.

Two threads, Kate. First, this was the closed-versus-open inflection point, live. In one twenty-four-hour window, the U.S. government pulled the two most capable closed American models offline, and a Chinese lab shipped a frontier-class open model under an MIT license with the explicit message: this can never be switched off. Whatever you think of the security rationale, the strategic effect was to hand the open-weight ecosystem — much of it now flying a Chinese flag — its single strongest argument ever. The worry that writes itself is that the West regulates itself into dependence on an adversary's give-it-all-away play.

And the second thread?

Governments stopped debating AI and started reaching into the products. A White House export letter. Forty-two state AGs on the eve of an IPO. A UK evidence-fabrication probe. Three governments, three mechanisms, one weekend — all touching live products, not abstract policy. And the kicker, reported but worth flagging as unconfirmed: Altman, Amodei and Hassabis are said to be appearing together before world leaders at the G7 in France this week, June fifteenth to seventeenth. Days after Washington switched off Amodei's best model.

That is an awkward dinner table.

The most awkward, Kate. And the read I'd leave people with is uneasy but not defeatist. Fable is an American model, leading the world at the very thing it was punished for, switched off on the word of its own investor. The libertarian instinct here is loud — a power that easy to grant and that hard to claw back should make everyone nervous, regardless of party. The open-source crowd is right that a revocable, gated model is a single point of failure. But they're dodging who funds the alternative, and they're dodging whose flag most of it flies. The West can still win this. You just don't win it by quietly killing your best product on a Friday and hoping nobody asks to see the paperwork.

That's your AI in 15 for today. See you tomorrow.