On June 4, Anthropic published what read like a confession.

The blog post, co-authored by Anthropic Institute’s Marina Favaro and Jack Clark, warned that AI systems were approaching a threshold — recursive self-improvement, the point at which models begin building their own successors without meaningful human intervention. The authors did not say this was inevitable. They said it could come “sooner than most institutions are prepared for.”¹

The language was careful. The concern was genuine. And the request was extraordinary for a company at the frontier: that the world consider a coordinated slowdown, the safety research and governance structures be given room to catch up, that the option to pause remain open. It was, by any measure, a serious document.

It was also published just three days after Anthropic confidentially filed for an IPO — at a reported valuation near $965 billion.²

We are not suggesting the concern was fabricated. We are suggesting something more uncomfortable: that a warning issued from within a racing system takes on a different shape than a warning issued from outside it. It becomes part of the race. It becomes, in its own way, a signal — not to slow down, but to be seen as the kind of company that knows it ought to.

Five days later, Anthropic released Claude Fable 5.

The launch was not reckless. Before release, Anthropic ran over a thousand hours of internal re-teaming. External organizations attempted to find universal jailbreaks and failed.³ The model’s safeguards, Anthropic reported, activated in fewer than five percent of sessions.⁴ In high-risk areas — cybersecurity, biology, chemistry — responses were rerouted to a less capable model.

Every safeguard held. Every benchmark cleared. This is the part that matters: Anthropic’s safety team did not cut corners. The engineers asked the right questions — the questions they were given. No more, and no less. Whether the model could be weaponized. Whether its safeguards could be broken. Whether it was ready, by every technical measure they had designed.

What the tests did not ask — and what no classifier can — is “Why June 9?” Why five days after the warning? Why nine days after the IPO filing?

Those questions were not in scope. They were never going to be in scope. Because the boundaries of the test are set by the safety team; they are set by the system the safety team operates within. And that system had already answered the question of timing — on June 1, in a confidential document filed with the Securities and Exchange Commission.

The warning was real. The warranty was also real. They were issued by the same company, in the same week, from the same hand. This is not hypocrisy. It is something harder to name — and harder to fix.

On June 13, the U.S. Department of Commerce issued an export control directive to Anthropic.

The order was unprecedented: suspend all access to Fable 5 and Mythos 5 for any foreign national — whether inside or outside the United States — including Anthropic’s own employees.

To ensure compliance, Anthropic did the only thing it could. It shut the models down for everyone.

Anthropic disputed the order. The company said it had received only verbal notice of a “potential narrow, non-universal jailbreak”⁵ — one that, it argued, could be used to elicit similar capabilities from other publicly available models not subject to the same controls. “We disagree,” Anthropic wrote, “that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people.”⁶

They were not wrong to say so. But being right about the jailbreak was beside the point. Regardless, the models were down. Not for Anthropic, but for the people who had been using them.

This is what Charles Perrow called a normal accident — not a failure of any single part, but an outcome produced by the normal functioning of a complex, tightly coupled system. The safety team did their job. The IPO process followed its logic. The government acted on tis own calculus. Each part worked exactly as designed.

The system produced this result not despite functioning correctly, but becuase of it. This was not a failure. This was the system working as designed. And that is the problem.

Anthropic warned us. They also warranted us. These are not contradictions — they are the same gesture, made by the same hand, in the same week.

A warning issued by those who cannot afford to stop is not a warning; it is a record. It says: we saw this coming. It does not say: we chose to wait.

The tests passed. The model launched. The access was cut. And somewhere in that sequence — between the IPO filing, the safety blog, and the thousand hours of red-teaming — the question that mattered most was never asked.

Not: Can this model be misused? But: Should this model exist in a world not yet ready to govern it?

That question has no classifier. It has no benchmark. It cannot be red-teamed.

It can only be asked by those with nothing to prove — and nothing to sell.

Polanyi is an independent index of what technology does to us. This is the first letter.