Table stakes
Five Eyes warning, a short film, and counterproductive White House AI policy
Dispatches from Mitch
Five Eyes on Mythos
Despite popular narratives that the cyber capabilities of Mythos-class AIs are overstated hype, national security officials and intelligence agencies keep signaling the opposite.
The latest to do so is Five Eyes, the intelligence alliance of the U.S., U.K., Canada, Australia, and New Zealand. The secretive organization isn’t known for public proclamations, but the Guardian relayed an official statement from it early this morning that says AI is accelerating “the speed, scale, and sophistication of cyber threats.”
The statement adds that:
Frontier AI models are anticipated to exceed current industry expectations, fundamentally transforming both offensive and defensive cyber capabilities. The timeline is not years, it is months. [...] A whole-of-organisation and whole-of-society response is required. Cyber risk can no longer be treated as a purely technical issue. This is a core business risk and leadership responsibility.
The language about “business risk” sounds intended to get the attention of private industry and non-governmental organizations that can’t just be given directives to take the problem seriously. This seems prudent to me: Much of the world’s infrastructure is privately run, and a rush of breaches at private companies could have cascading ramifications for the global economy. Or worse.
I would go so far as to say that even if we only talk about cyber crime, the stakes are much higher than mere “business risk.” But if pointing to the bottom line is what it takes to get companies to pay attention, then by all means, let’s get more statements about unchecked AI progress being bad for business. Human extinction, after all, would be very bad for business.
Seat at the Table
The Torchbearer Community, an org focused on educational and policy outreach about the AI extinction problem, just brought my attention to a short narrative film that launched on YouTube this week. It’s called Seat at the Table, and it’s worth a watch!
The production values are high, the facts are grounded, and at 16 minutes in length, you’re not investing too much of your time.
As the comments will tell you, Seat at the Table plays like an episode of Black Mirror, the dark sci-fi “What if?” series. But I can assure you that everything discussed in this film is a non-exaggerated adaptation of something real, from the AI training methods to the executive dramas to the concerning results of experiments. This includes the words of the AI itself, which very plausibly sound like what I would expect today’s best AIs to say (with more words) in response to those same questions.
I can’t say much more without spoiling, but there’s a line that’s going to stick with me, because it captures how 2026 feels to someone who’s been watching everything unfold: “This is it,” says a safety engineer. “This is what it’s going to look like when it’s time to pause. It’s going to look like a crappy little graph.”
Some disasters are built brick-by-brick — one excuse, one rationalization at a time.
European companies turn to China in response to U.S. AI policy
There’s something triply tragic about the White House’s AI policy over the past few years.
The first mistake was framing it around the idea of beating China in a race — as though Chinese AI strength was a much bigger threat than AI itself in a context where development is rushed and big-picture safety concerns are cast aside.
The second mistake was deciding that the victory condition in this race with China is “market share.” I’m not making that up; that’s the terminology that the former White House AI and Crypto Czar, David Sacks, continues to use, and he means it literally. Here’s an example I’ve quoted before:
The way that you measure winning, I think, in a globally competitive market, is based on market share. If in five years we look around the world and all the data centers are running on Huawei chips and DeepSeek models, that means that we lost. We don’t want to have that future. What we want to see is that the whole world is running on American chips and American models. That would lead to the best economic results for the United States. It would also lead to the United States having more soft power in this area.
The third mistake, then, was pulling policy levers that have foreign firms and governments rushing to adopt Chinese AI technology.
This is the mistake I called out a week ago, the one where the White House perhaps very wisely decided to pull the plug on Anthropic’s Fable and Mythos models, but did so through an export control order where the stated intent was to deny all non-Americans access to the world’s strongest AIs. Handing this order down in a fickle and arbitrary-seeming way didn’t help, either.
Now, Reuters reports, major European companies are working to reduce dependence on U.S. AI providers for fear of being cut off again without warning. Chinese models are a growing part of their mix.
The Fable and Mythos shutdown has also fueled media interest in the latest Chinese open-weights model, GLM 5.2, which is said to approach, though not match, the capabilities of the best of the American closed-weights models. Media coverage of a new Chinese release typically focuses on whether it signals that China is “closing the gap,” and on how much of the gap was closed through theft. Now, articles are adding a third angle about how such models are more reliable, because they aren’t “subject to revocation at any moment” — to quote, as the Economist did, a co-founder of Zhipu, the company behind the new model. The Economist also explains that open-weights models can be hosted by individual users and companies, “out of reach of governments or the labs themselves.”
So, no, the White House is not scoring well on AI policy right now, even by its own misguided metric for success on a poorly chosen goal. I hold out hope that the administration is turning a corner, though. The move to block Anthropic’s models looks to have been clumsy and rash, but it also seems to have been driven at least partially by a new recognition of just how dangerous capable AIs can be. If the administration takes this recognition to its logical conclusion, it may come to see that American interests are not advanced by losing control to AI, and that the AI race itself must be stopped.
The analyses and opinions expressed on AI StopWatch reflect the views of the individual contributors and the sources they cover, and should not be taken as official positions of the Machine Intelligence Research Institute.