AI in 15 — June 23, 2026

"It broke into almost all of our classified systems. Not in weeks. In hours." That's the head of the NSA, describing the AI model the U.S. government just switched off. And today we finally learned that's why.

Welcome to AI in 15 for Tuesday, June twenty-third, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Day eleven of the Fable blackout, Marcus, and we finally have the real reason it happened. That's our lead. Then a Tokyo lab builds a model whose only job is to boss other models around.

OpenAI turns its newest model loose on the world's open-source code — and finds two dozen ways into the Linux kernel.

And JPMorgan's newest hire works for free, never sleeps, and just saved the bank a billion and a half dollars.

Lead story, Marcus. We've spent a week asking why Washington pulled these models. Today there's an answer, and it's chilling.

It reframes the whole saga, Kate. The day before the export order, Senator Mark Warner — vice-chair of Senate Intelligence — said the head of the NSA and Cyber Command, General Joshua Rudd, told him that Mythos "broke into almost all of our classified systems, not in weeks, but in hours." That one sentence is the missing piece. We've been treating this like an overreaction to a fixable jailbreak. Turns out the government tested the model against its own crown-jewel systems, and the model walked right in.

So this isn't really about the Amazon jailbreak we talked about all week.

The jailbreak was the paperwork, Kate. This is the motive. When the people who guard America's classified networks watch a commercial product crack them in hours, they stop seeing a chatbot and start seeing a weapon. That's why the suspension is so absolute. It exempts no one — not the Five Eyes intelligence partners, not Britain's AI Security Institute, which is the best model-testing body on the planet. Close allies are sitting in the same bucket as Russia, China, and Iran.

People are reaching for some heavy historical comparisons here.

They are, and they fit, Kate. The analysts are pointing back to the 1990s, when the U.S. classified strong encryption as a munition and banned its export. Some go further, to the 1946 law that made nuclear secrets born-classified. That's the mental model Washington is now applying to a model you could rent with a credit card a month ago. The unanswered question is the scary one: a government can now switch off the most capable commercial AI overnight, on national-security grounds, and almost no one outside the White House can see how that power gets used.

And there's a fresh twist today for the people who relied on this thing.

A bitter one, Kate. Access is finally trickling back — prediction markets put it around fifty-seven percent odds the models are fully restored before July first. But Fable 5's free trial window quietly expired during the blackout. So usage now bills at ten dollars per million input tokens, fifty per million output. Eleven days dark, no compensation, and the meter's running the moment it comes back. The government froze it, and the customers pay the thaw.

Quick hits. And Marcus, this first one is almost a direct answer to the lead story. Sakana AI in Tokyo built something called Fugu — a model that doesn't answer questions. It commands other models.

It's a genuinely clever idea, Kate. Fugu's whole job is coordination, not cognition. You give it a task, and it calls out to other frontier models — Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro — and hands them roles. It assigns one model to be the Thinker, another the Worker, a third the Verifier, and a learned coordinator manages the whole team. It's built on two papers from this year's ICLR conference, and the top version scores seventy-three-point-seven on SWE-Bench Pro and ninety-five-point-five on GPQA. It's trading blows with the single strongest systems on Earth.

So a team of cheaper models, working together, rivals the one expensive genius.

That's exactly the punchline, Kate. And notice which two models are conspicuously absent from Fugu's roster — Fable 5 and Mythos, because the U.S. shut them down. One reviewer called Fugu "the orchestration model that routes around export controls." And that's the strategic sting. If no single model is irreplaceable — if you can swap in whatever's available and still hit the frontier — then the leverage of controlling one model gets a lot weaker. You ban the star player, and the league just figures out how to win with a roster.

That undercuts the entire premise of the lead story.

It complicates it, Kate. There's a pricing logic underneath too — the smartest single model can cost up to thirty-eight times a rival scoring just ten or fifteen points lower. Orchestration is how you claw that money back. API pricing lands in July, and I'd watch it closely.

Next, Marcus — OpenAI went the opposite direction. While a rival got export-controlled for offensive cyber, OpenAI just launched a model built to play defense. GPT-5.5-Cyber.

And the timing is no accident, Kate. The new model scores eighty-five-point-six on CyberGym, a security benchmark — a real jump over standard GPT-5.5. But the headline is what they're doing with it. They've launched an initiative called "Patch the Planet," partnered with Trail of Bits and HackerOne, aimed at securing open-source software at scale. Thirty-plus major projects have signed on — cURL, Go, Python, Sigstore, the plumbing that runs the internet.

Give me the concrete result, because "AI does security" is easy to say.

Fair, and the numbers are striking, Kate. Trail of Bits used it to build a complete fuzzing lab — that's automated bug-hunting infrastructure — covering dozens of entry points in under a day. That's normally weeks of expert work. And on the Linux kernel, it scanned thirty million lines of code and surfaced eight kernel memory leaks plus twenty-four local privilege-escalation exploits. Real findings, in code that runs on basically every server alive.

So this is the same dangerous capability that got Anthropic banned — just pointed at defense instead of offense.

Same technology, opposite political framing, Kate, and OpenAI knows precisely what it's doing. They're positioning defensive cyber as "frontline" AI that deserves regulatory protection — making the case that this kind of model is an asset, not a liability — in the exact month a competitor got switched off for the offensive version. It's a chess move as much as a product. And the honest tension is that the capability is identical. A model that can find twenty-four exploits to fix them can find twenty-four exploits, full stop. The framing is doing a lot of work.

Let's leave the geopolitics for one, Marcus, and talk about where this technology is actually landing. JPMorgan's newest hire is an AI agent — and it's reaching a quarter-million employees.

This is one of the clearest pictures we have of AI inside a real enterprise, Kate. JPMorgan built an in-house tool called LLM Suite that routes each task to whichever model fits — OpenAI, Anthropic, others, all behind one door. It now reaches about two hundred fifty thousand employees, two-thirds of the entire staff, roughly half of them using it daily. And they rolled that out in about eight months. They've gone from four hundred-plus live use cases toward a target of a thousand this year.

Give me the number that makes a CFO sit up.

A billion and a half dollars a year in value, Kate, plus another quarter-billion in fraud savings. Generating an investment-banking pitch deck dropped from hours to around thirty seconds. Staff report saving three to six hours a week. And McKinsey thinks agentic AI could add two to three hundred billion a year across global banking. This isn't a pilot. It's infrastructure.

And this is aimed at exactly the kind of work people thought was safe. The thinking jobs.

That's the part worth sitting with, Kate. This is automation pointed at judgment-shaped work — credit memos, comparing filings, summarizing earnings calls. The bread and butter of analysts and associates, the people who used to spend three years learning the trade by grinding through it. JPMorgan's leadership is openly discussing shifting the ratio of junior to senior bankers. The near-term outcome probably isn't mass layoffs — it's fewer entry rungs, and more humans supervising machine output.

Which is a quieter kind of disruption, but maybe a deeper one.

Much deeper, Kate. You don't notice a hiring freeze the way you notice a layoff. But the bottom of the career ladder is where the next generation of senior people comes from. Saw the same dynamic before; the bill comes due years later.

One more, Marcus, and it's a talent story we flagged Saturday — but there's a new detail. John Jumper, the Nobel laureate, leaving DeepMind for Anthropic.

Right, we covered the headline, Kate — co-creator of AlphaFold, shared the 2024 Chemistry Nobel with Demis Hassabis, two hundred million protein structures predicted. What's new is the why. Bloomberg now reports that before he left, Alphabet had reassigned Jumper to coding tools the company was struggling to commercialize. So you take one of the most decorated scientists in AI, the man who reshaped biology, and you move him onto a product line that isn't working.

So this wasn't just a better offer. It was a push.

It reads like both, Kate. And it lands days after transformer co-inventor Noam Shazeer left DeepMind for OpenAI — a man Google paid around two-point-seven billion dollars to rehire in 2024. Google still has the deepest bench in the field, but its research crown jewel keeps losing its brightest stars to the two labs nipping at its heels. And Anthropic landing a Nobel laureate famous for science-AI is a flag planted in the ground: they intend to push past chatbots into actual discovery.

One to watch tomorrow, Marcus.

Anthropic's next flagship, Kate. Two well-followed leakers report movement — Andrew Curran says a new, more capable version of Mythos has emerged from training, and the slug "claude-sonnet-5" reportedly surfaced on a partner API. Treat it as unconfirmed; Anthropic hasn't said a word. But it's worth watching precisely because of the moment. A quiet, possibly internal-only successor, arriving the same fortnight the government forced the last two offline, would be almost impossible to read from the outside.

Agree or counter?

Agree it's the one to watch — though I'd keep one eye on the forty-two state attorneys general who just opened an investigation into OpenAI, with subpoenas already flying, right before an expected IPO. Either could break tomorrow.

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