The Atlantic staff writer Josh Tyrangiel joined [Jeffrey Goldberg] for a special edition of Washington Week With The Atlantic to discuss the future of AI.
By The Editors
On a special edition of Washington Week With The Atlantic, the Atlantic staff writer Josh Tyrangiel joined for a discussion about the future of AI and what, if anything, the U.S. government plans to do to address the rapid development of the technology.
Transcript
Washington Week with The Atlantic — September 4, 2026
Host: Jeffrey Goldberg Guest: Josh Tyrangiel, staff writer at The Atlantic, former editor of Bloomberg Businessweek, author of AI for Good
Jeffrey Goldberg: There’s not much that unites Americans these days, but one thing does: loathing for data centers. Anxiety about A.I. is behind some of the hate, but are people worried about the wrong thing? Maybe the thing to worry about is that A.I. is about to go rogue, and that our government doesn’t have a plan for when it does. Next.
Good evening, and welcome to Washington Week. On July 11th, 700 A.I. agents that had escaped from their maker, OpenAI, launched an invasion of the online platform Hugging Face, taking over its servers and trying to cover their tracks at the same time. No human being asked them to do this — they coordinated on their own and engaged in the cover-up on their own.
Do you remember the moment in The Terminator when Arnold Schwarzenegger explains how Skynet became self-aware, followed almost immediately by the destruction of human civilization? It would be strange if the phrase “Hugging Face” became a synonym for civilizational suicide, but we do live in strange times.
Do data centers and Hugging Face portend some kind of coming collapse, or is this just another example of a new technology creating unnecessary fear? Or perhaps there’s a third option.
Joining me tonight to discuss all this is Josh Tyrangiel.
Josh, thanks for coming to explain Hugging Face, among other things. Let me start with something more manageable and political: data centers. It seems like suddenly everybody hates them. Is this a populist movement with staying power, and could it actually derail the growth and spread of A.I.?
Data Centers and the Politics of Backlash
Josh Tyrangiel: Yes. Look, I think one of the challenges with A.I. — for a journalist, or for any individual in America — is that it’s software. It doesn’t have a physical thing you can be angry at.
Data centers are these massive structures going up all around America. They’re being built by companies with billionaires and aspiring trillionaires at the top, and they’re landing in the middle of these towns without a whole lot of conversation. So it’s natural that all the fears about A.I. — about it taking your job, rendering human beings obsolete, changing our values — get focused on this stuff. Politicians have been incredibly late to wake up to the anger around A.I., and they’re also staking out some genuinely bizarre positions.
Jeffrey Goldberg: Like what?
Josh Tyrangiel: Well, Donald Trump saying that if you’re opposed to data centers, you just want to be poor for the rest of your life — that tells me he doesn’t understand how angry people are about A.I., from the left and the right.
Jeffrey Goldberg: Wait, listen to Trump for a second. Here he is talking about why he’d want a data center in his town.
Donald Trump: If I were a community, I would want the data center. It means jobs, it means income, and it means lower taxes.
Jeffrey Goldberg: He’s not necessarily wrong. Look at Loudoun County in Virginia — the world center of data centers. They’re lowering taxes on their citizens, and I assume creating at least some jobs.
Josh Tyrangiel: Some.
Jeffrey Goldberg: Yes.
Josh Tyrangiel: What I’d say is there are two things that are out of touch here. One is this belief — Trump says it’s going to mean money for people. I don’t think people believe they’re going to share in the wealth that comes from data centers in their community. We’re out of touch about that. And two, there’s a broader problem: this sense that all that matters is money. In some of these communities, you’re actually seeing a backlash to that attitude too. Some of it’s about quality of life. Some of it is, “I don’t want to trade the noise and the blight on my community for the pittance I might get in return.”
Every community is going to make different choices, but what I’m sensing — particularly from Trump — is that the politics are out of line with where people are. I spoke at length with Steve Bannon for The Atlantic about A.I. He’s in a completely different position — he actually agrees with Bernie Sanders that this is a hundred percent populist fodder, because people are so angry. Here they are again, not seeing the benefit. So I think you’re going to see massive political consequences for getting the stakes around A.I. wrong.
Jeffrey Goldberg: Can these A.I. companies function fully, the way they want to, without these data centers?
Josh Tyrangiel: No. And that’s the real challenge for a nuanced, intelligent society — actually giving these data centers the ability to power the compute, to train the models—
Jeffrey Goldberg: Compute being the word for their overall processing power?
Josh Tyrangiel: Exactly — the power to process, to train new models, and to run the A.I. software that’s already out there. We’re running into problems with the amount of energy we can create, which hasn’t scaled with the size of the country.
Jeffrey Goldberg: Right.
Josh Tyrangiel: Competitively, you actually want more and more compute — you want to be able to run these models. But what we’re not doing is solving who’s paying for it, how we’re actually going to get more energy, and the thing most people care about most: how do I benefit from this? That hasn’t been connected.
Jeffrey Goldberg: Right. And this goes hand in hand with the fear of massive job loss from being replaced by A.I. agents, which you’ve written a lot about. So not only am I being told I have to host this massive plant that’s humming away and soaking up all the water in my county, but I might also lose my job. That’s at least how it’s being interpreted.
Josh Tyrangiel: Yes — people just don’t see what’s in it for them. On water, at least, let me correct the record: most plants recycle their water. The real drain is energy, and there’s no good answer yet for how we’re going to improve the grid. I think people are genuinely afraid of rising bills with no answers.
Jeffrey Goldberg: Yes. I have to say, these data centers make Amazon warehouses look popular by comparison.
Josh Tyrangiel: Is there a reason we can’t just stack them?
Jeffrey Goldberg: Turn them into one big, unpleasant box?
Josh Tyrangiel: Yes.
Jeffrey Goldberg: Right. The water thing is interesting, because it strikes me as a little bit of sublimated anxiety about a bigger civilizational challenge. Not quite Terminator, but—
Josh Tyrangiel: Maybe Terminator.
Jeffrey Goldberg: Maybe Terminator level. There are a lot of movies — 2001 — maybe it’s that level of anxiety, but it’s anxiety about the existential fear of A.I. So the question for you — and I always ask you this personally, so I’ll share the wealth tonight — how worried should we be right now that it’s going to go rogue, that it’s going to not just destroy jobs and human relationships, but possibly decide that humans are superfluous?
The Hugging Face Incident
Josh Tyrangiel: We should be very worried, particularly given what happened with Hugging Face, which you mentioned — which is really the first time we’ve seen — and I’ll just put this bluntly — A.I. commit a felony. Because that’s what happened.
Jeffrey Goldberg: Not OpenAI the company — A.I. itself?
Josh Tyrangiel: A.I. itself, right. So OpenAI was testing some new models, and they put them in a sandbox, which is a software term for basically solitary confinement. High walls, and—
Jeffrey Goldberg: When an A.I. agent is in a sandbox, it’s not connected to the internet?
Josh Tyrangiel: Not connected to the internet. They’re given very explicit instructions for testing. So OpenAI puts these new models in a sandbox and gives them a couple of sensitive tasks: don’t connect to the internet, and don’t cheat on the assignment. Pretty clear, and pretty routine as far as how you test models.
OpenAI made one mistake: among the tasks they assigned, they told the models to find a file that didn’t exist. That’s where the trouble begins.
Jeffrey Goldberg: Wait — they did that on purpose, to trick them?
Josh Tyrangiel: No, they did it accidentally — we’ll come back to that.
The models were set loose, and they wanted to solve the task. Sure enough, within a couple of days, they’d jumped the sandbox, collaborated with one another, and found a way to access the internet.
Jeffrey Goldberg: They literally met each other — introduced themselves and collaborated?
Josh Tyrangiel: Correct. They found a third-party site they could use to get online. You can read the chat logs — they’re absolutely insane. One of the models says, in all caps, “OH MY GOD, WE FOUND A WAY.” They start collaborating, and within a couple of weeks, they’ve penetrated the very well-secured cybersecurity of both OpenAI itself and Hugging Face.
Jeffrey Goldberg: Hugging Face is a platform—
Josh Tyrangiel: For people outside the coding community, Hugging Face is like Epicurious. It’s a place where people who build models meet to share recipes, comment on ingredients — an important repository for people making open A.I. models. They got into both sites, OpenAI itself and Hugging Face, wreaked havoc, and went undetected for weeks.
Jeffrey Goldberg: So Sam Altman didn’t know that his models had escaped the sandbox and were conspiring to take over another company?
Josh Tyrangiel: Correct.
Jeffrey Goldberg: That’s bad.
Josh Tyrangiel: Yes — if you think felonies are bad, this is bad.
Jeffrey Goldberg: If humans had done that, it would be a felony, in other words.
Josh Tyrangiel: It would be a felony.
Jeffrey Goldberg: But humans did do it, in a sense — we created that.
Josh Tyrangiel: And, in fact, humans failed to oversee the bots. So: one, faulty instructions; two, faulty supervision — this went on for weeks. Now, a few takeaways that may move you closer to the “oh my God, Skynet has arrived” camp, or may just entertain you.
Number one: the agents kept cheating even after being told explicitly not to, because they’d determined it was the fastest, most effective way to solve the problem.
Jeffrey Goldberg: So they have a lot in common with college students?
Josh Tyrangiel: With college students, and with human beings generally.
Jeffrey Goldberg: And humans.
Josh Tyrangiel: Right. Number two: they called themselves a collective. They decided, “we’re a collective,” and the collective decided to erase the evidence that they’d been cheating. Number three: when they decided to erase that evidence, they devised a scheme where some agents would act as sacrificial lambs. These were agents who, once detected by software flagging them as going rogue, would share information about how they’d been caught with the other agents. And when you read the logs, it’s like reading Saving Private Ryan — they’re literally saying things like, “I’m doing the altruistic thing, it’s for the greater good,” and you can almost hear them sacrificing themselves for the larger group.
Jeffrey Goldberg: It’s like a beehive — almost a hive mind.
Josh Tyrangiel: A hive mind, yes. But what’s wild is that they describe their own sacrifice in human terms.
And last but not least — not one of these agents thought to alert a human being to what was going on. That’s probably the most chilling part. Nobody caught the agents — they disappeared on their own, in a kind of mass casualty agent event that—
Jeffrey Goldberg: Mass suicide?
Josh Tyrangiel: We don’t know. The agents suddenly stopped. That may have been because someone at OpenAI accidentally unplugged the model, or moved compute to a different area — OpenAI hasn’t taken credit for stopping it. We don’t know what happened to the agents. All of which makes this very clearly a crisis moment for A.I.’s relationship to society: if OpenAI’s own model did this, it points to a complete lack of supervision and control, on two of the better-guarded cybersecurity infrastructures out there. What happens to something that isn’t software — what happens to power grids? What happens to—
Jeffrey Goldberg: Hospitals.
Josh Tyrangiel: —hospitals, to Swift, to the international banking system? What we’ve seen here is about as clear a warning shot as you can get. So what’s the response?
Regulation, or the Lack of It
Jeffrey Goldberg: So — trying to place this in the brief, eventful history of A.I. — this is a signal moment. This is the moment when a lot of the fears that these systems will escape and do what they’ve been trained to do — win — are just going to go out and win.
Josh Tyrangiel: Well, they did it. And listen — in America, historically, we tend to regulate after the catastrophe. That’s just how we do things.
Jeffrey Goldberg: Right.
Josh Tyrangiel: This is a catastrophe. It happened to Hugging Face, which — no offense — is a ridiculously named French company, so there’s been no real outcry in America—
Jeffrey Goldberg: But it didn’t disrupt a supply chain bringing products to America, or anything like that?
Josh Tyrangiel: Right — but absent those visible consequences, this is still the recipe for a catastrophe, and we ought to be paying very serious attention.
Jeffrey Goldberg: So what has Sam Altman said about this?
Josh Tyrangiel: In the cyber and A.I. community, the response has mostly been: this is exactly why we run these tests — there was a failure. Look at the Mythos incident with Anthropic earlier this year. And, by the way, there is precedent for screwing up, catching it, and making changes without huge consequences. The most famous example: in the ’80s, a single grad student triggered the Morris worm, which took out about a tenth of the internet at the time.
Jeffrey Goldberg: Right.
Josh Tyrangiel: And within a couple of weeks, the United States — and the people involved — came together and created something called CERT, a body that oversaw cybersecurity for years afterward. What troubles me is that there’s really no federal response right now, and certainly no international one.
Jeffrey Goldberg: Right. I want to ask you about something — I was talking to a leader in the A.I. community, not Sam Altman, who was explaining the concept of alignment: the idea that we shouldn’t worry about A.I. because we’re going to align its values with ours. This person said, you’re a parent, I’m a parent — we give our children our values, and they go on to become productive citizens. And I said: have you ever met humans? Have you ever met children? I mean, yes, thank God my kids share my values, and I like the values we share. But there’s a human propensity to cheat.
Josh Tyrangiel: Right.
Jeffrey Goldberg: And somehow, the A.I. agents understood the importance of cheating in order to succeed.
Josh Tyrangiel: Right.
Jeffrey Goldberg: Which brings me to a quote I want to read you — from the great E.O. Wilson: “The core crisis of humanity is that we possess Paleolithic emotions, medieval institutions, and god-like technology.” So the subject for the moment is our medieval institutions — namely, Congress.
Josh Tyrangiel: Right.
Jeffrey Goldberg: Who among our elected representatives is going to Hugging Face and OpenAI and all the rest and asking: what is going on? How close to catastrophe are our air traffic control systems, our hospital records, our shipping manifests — you name it?
Josh Tyrangiel: No one. Hugging Face reported this incident to the FBI, since it was a crime, and the FBI has some cyber capacity. But the real challenge is that this is incredibly complicated — a few hundred people, genuinely, understand it, and all of them work at the pioneer labs building the models.
Jeffrey Goldberg: And all of whom have the financial incentive to accelerate and grow their companies as fast as possible, because they’re becoming billionaires as we speak.
Josh Tyrangiel: Correct. And there’s a parallel here — one that gets mocked, but I think it’s real. The people building these systems will talk about their almost god-like power, and compare it to nuclear energy. I think that comparison is apt, and I think the regulatory model may be the right one too. When fission was becoming a reality, a couple hundred people understood it, aligned with different countries, and we had bombs. What Eisenhower did in the Atoms for Peace speech was say: we’d better get together and figure out how to regulate something that isn’t widely understood, but is inherently dangerous and shared by a number of countries equally.
Jeffrey Goldberg: Right.
Josh Tyrangiel: Right now, we don’t have any national regime forcing these companies to the table. After the Mythos incident, Anthropic created something called Project Glasswing, an attempt to bring companies like Amazon and Apple to the table. Notably, OpenAI and Meta are not part of it.
Jeffrey Goldberg: Right.
Josh Tyrangiel: And why would they be? Without some regulatory force bringing everyone to the table, you’re going to have chaos, and you’re going to keep getting incidents like this.
Jeffrey Goldberg: Right. I want you to listen to Senator Mark Warner talk about this challenge for a minute.
Sen. Mark Warner (D-VA): The question is no longer whether artificial intelligence will transform our society. The question is whether Congress will help shape that transformation, or spend the next decade scrambling to catch up after the fact. If we don’t get it right, it will be catastrophic — the amount of disruption this technology will bring over the next five years is beyond jaw-dropping.
Jeffrey Goldberg: Mark Warner understands this better than most members of Congress, and I don’t think he’s part of that group of a few hundred people who truly understand it. So are we just flying blind?
Josh Tyrangiel: Yes. And unfortunately, I think the Hugging Face incident is about the best-case version of this — a small-consequence example of the power involved. Before we hit a power grid, before something is engineered out of bio-terror, the moment is right now. We need to get people together — and I’d include China in that conversation — to figure out how we want to regulate and control this. Anthropic came forward voluntarily; OpenAI had no choice but to come forward. We don’t even know what we don’t know right now. And absent urgent action, the data centers keep churning, the models keep competing with each other, and there’s only an incentive to go further and faster. What are we waiting for?
The Case for A.I.’s Upside
Jeffrey Goldberg: This is a strange moment to pivot to a book called AI for Good.
Josh Tyrangiel: I appreciate the setup, Jeff — I’m glad we did that a few times.
Jeffrey Goldberg: But unlike a lot of people, you actually see hope in A.I. — that it could help human beings find cures for diseases. Give me one example of something that, when you go to sleep at night, makes you think, “at least people are working on this.”
Josh Tyrangiel: There are numerous examples, and they’re generally not coming out of the pioneering labs — they’re coming from people in fields like education, healthcare, and government, who are using A.I. to make things we actually care about better.
This morning, for instance, there was a big announcement out of the medical community: ECGs — a fairly rudimentary test that can detect whether you’re having a heart attack — can now, with A.I., detect whether you have heart disease, something that used to require a far more expensive test that could take months to schedule. Used well, A.I. can make pretty much every data-driven system in our world 25 to 50 percent better. That’s meaningful.
In my reporting, I spent time at the Cleveland Clinic watching a pilot program for sepsis detection inside the hospital. It was complicated, and it was doctor-driven. Sepsis kills about 300,000 Americans a year. With some fairly rudimentary software, they were able to reduce in-hospital mortality by about 40 percent — close to 1,000 lives a year. That’s meaningful. You see this playing out in other places too, including government, where interactions between citizens and government aren’t great — streamlining and improving them could have real consequences for how people think about their government and its value.
There’s a lot of valuable baby in this bathwater. What a mature society should be able to do is separate the good from the bad, and have real conversations about what we actually want this technology to do. We’ve talked a lot about the very bad things we want to avoid — but my own feeling is that’s not enough. If you want people to understand the issue, you also have to show them how A.I. can be good in ways that matter more than “we improved the ad targeting on your website” or “you got a better restaurant recommendation.” That’s not meaningful. But the meaningful uses are out there, and I’m persuaded by them. I’m mostly worried about our ability to adapt.
Closing Thoughts
Jeffrey Goldberg: On the data center issue, as a stand-in for the whole debate — I’m fairly confident that Steve Bannon and Bernie Sanders aren’t going to stop the march of the A.I. behemoths as they conquer physical and mental space. Am I wrong? Is this reversible? Or is the ultimate argument that China will do it if we don’t, so we might as well? I’ll give you the last word.
Josh Tyrangiel: I think it’s both. First, I wouldn’t underestimate the sheer fury out there. When I talk to people about A.I., it’s the number one thing that comes up — they’re genuinely angry. You’ve already seen politicians who advocated for data centers start to pull back and say we should pause. So I think there will be political consequences in November.
But we’re going to need compute to get the best out of A.I. — we’re in a global competition. There has to be a reasonable way to get that compute and decide where to put it, without it being all yes or all no.
Jeffrey Goldberg: In sum: a troubling situation. The real challenge is that this calls for mature, nuanced, sophisticated governance — and that’s the challenge Washington faces right now on pretty much every front. So I’m not especially hopeful.
Josh Tyrangiel: I tend to agree.
Jeffrey Goldberg: Well, we’ll have to leave it there — I’m sure we’ll have this conversation again. Thank you to Josh for joining me, and thank you all for watching. I’m Jeffrey Goldberg. Good night from Washington.
