Hello, dear readers!

This week, one story ate the newsletter.

It started with a former Anthropic researcher warning that AI could kill us all. Then a PR firm appeared. Then we started wondering who benefits when the public becomes sufficiently terrified of frontier AI. And somehow we ended up back in the Cold War.

So today: AI doom, regulatory moats, and why mutually assured destruction may actually be the reassuring comparison.

Remember Jacob Coxon?

Earlier this month, the former OpenAI and Anthropic researcher quit his job and posted a spectacular warning: people building frontier AI genuinely believe it could kill everyone before the end of the decade. “This is not a marketing stunt,” he added.

The post went enormously viral, Coxon landed interviews across major US media, and his resignation became one of the biggest AI-safety stories of the year.

Then came a slightly awkward follow-up. Pirate Wires reported that DEY. Ideas + Influence, a communications firm working with prominent figures in the AI-safety movement, had been helping arrange media appearances for Coxon. The evidence does not show that DEY engineered his original post: the email obtained by Pirate Wires came the day after it was published. Coxon may sincerely believe every word he said, and getting PR help afterwards does not make his concerns false.

But it does make the episode more interesting. What looked like a spontaneous whistleblower moment became a professionally amplified media event remarkably quickly. And this particular media event was not selling sneakers. It was strengthening a political argument: frontier AI is extraordinarily dangerous, and somebody needs to constrain the people building it.

Which brings us to the people building it.

Anthropic itself openly supports tougher regulation of frontier AI. Its proposals include mandatory risk testing, independent evaluations, stronger security requirements and, for sufficiently dangerous systems, government power to block deployment.

There is a perfectly coherent argument here. If the most capable models pose the greatest potential risk, regulate the companies building the most capable models.

But the same rules look different from the other side. Extensive evaluations, security programs, reporting requirements and compliance teams cost money — and the companies best equipped to absorb those costs are precisely the giants already spending billions at the frontier.

A safety barrier can also be a competitive moat.

That creates an awkward split among people worried about AI. One group sees OpenAI, Anthropic and Google building increasingly powerful technology and concludes that governments need to constrain them. Another sees the same companies accumulating extraordinary power and worries that heavy regulation could lock that power in place.

The uncomfortable part is that both can be right at once. Regulation may genuinely reduce those risks, and the resulting rules may make it harder for anyone outside today's incumbents to compete.

This starts to resemble the old nuclear dilemma. A handful of actors possess an extraordinarily powerful technology, nobody wants to fall behind, and everyone has an incentive to keep building because everyone else is building too. Eventually, the argument goes, you need arms control.

But this is where the analogy breaks.

Nuclear arms control tried to stabilize an arsenal. If you froze the number of nuclear powers, you had at least constrained the spread of the weapon. With AI, freezing the players does not freeze the technology. Five dominant frontier labs can still keep making their models more capable behind increasingly expensive regulatory walls.

And AI has another property nuclear weapons conspicuously lack: it makes money.

The bomb itself does not attract hundreds of millions of users, sell subscriptions and API access, automate corporate work and generate the revenue needed to finance a better bomb next year. AI does.

So imagine a successful frontier-AI regulatory regime. Fewer companies can afford to compete, while those that remain keep improving and monetizing their systems. The barrier-to-entry effect is relatively easy to understand: the more expensive the requirements, the more capital you need to stay at the frontier.

What is much harder to know is whether the same regime actually prevents the catastrophe it is designed around. There is no control group for the singularity, no way to demonstrate in advance that a particular set of evaluations, reporting rules or deployment restrictions will stop a hypothetical runaway system.

And that leaves us with a curious asymmetry.

Nuclear arms control tried to keep the weapons from proliferating. AI regulation may end up keeping the companies from proliferating instead.

The moat is the part we can see. Whether it also works as a blast wall against the apocalypse is the part nobody can demonstrate in advance.

Thanks for reading AIport. Until next Monday — by then, the moat may be taller. Whether the apocalypse is any farther away is harder to audit.