Anthropic called a new model too dangerous to release, shipped a watered down version anyway, and then retired it within 72 hours under a US directive usually reserved for chips and weapons.
The phone call came on June 11. Andy Jassy, the chief executive of Amazon, dialed the US government — including Treasury Secretary Scott Bessent — to report that Anthropic's newest model had been jailbroken, and that the techniques for doing so were now circulating, according to The Wall Street Journal. Within 48 hours, Anthropic had retired the model for every customer, citing a federal directive. The model had been public for two days. "Kayfable" by Alberto Romero, The Algorithmic Bridge
To make sense of why that sequence mattered, you have to compress the previous six weeks into four beats. The first beat is April 7, when Anthropic announced a system it called Mythos Preview and said it was too dangerous to release. Mythos-class systems are a tier beyond the models Anthropic was willing to ship, capable in principle of material assistance with advanced cybersecurity, biology, and AI research. The April announcement was, in effect, a refusal: a public statement that the next rung of capability was on hold because the safety case had not been made. The Algorithmic Bridge: "Kayfable"
The second beat is June 9, when Anthropic released Fable 5. Fable 5 is a Mythos-class model that ships with guardrails: certain categories of prompts (cybersecurity, biology, distillation, and AI research) are routed to an inferior model. That routing is the safety move. The point of the release was to put the new capability in reach of customers while filtering the most sensitive queries. The source describes Fable 5 as Anthropic's best-performing system to date. The Algorithmic Bridge: "Kayfable"
The third beat is the June 11 pressure campaign. The US government responded to the Jassy call with a directive that forces a domestic company to stop serving a product to foreign nationals, including the company's own foreign employees. Anthropic confirmed the directive, citing "national security authorities," and said it received the order at 5:21pm ET. The mechanism is the same export-control infrastructure that limits the export of advanced semiconductors. A reader who has only ever heard "export control" in the context of chips and weapons is now looking at the same tool pointed at a chatbot. Anthropic statement, June 12, via anthropic.com The Wall Street Journal: "Amazon CEO's Talks With U.S. Officials Triggered Crackdown on Anthropic Models"
The reported positions diverge on the underlying motive. The US government's framing, as reported by The Wall Street Journal and Semafor, was that the jailbreak of Fable 5 had surfaced capabilities that could be repurposed by China. Anthropic's position, as stated in its public statement, is that the directive did not provide specific details of its national security concern, and that its review of the jailbreak technique found only "minor vulnerabilities" that "other publicly-available models are able to discover." Anthropic has denied that the China framing was the actual basis for the directive. The Wall Street Journal: "Amazon CEO's Talks" Semafor: "White House move to limit Anthropic linked to concerns about Chinese access to Mythos"
The fourth beat is June 12, when Anthropic retired Fable 5 for all customers — including foreign nationals and foreign-born Anthropic employees — followed on June 14 and 15 by meetings with the Trump administration, including Commerce Secretary Howard Lutnick and National Cyber Director Sean Cairncross, according to The Wall Street Journal and CNBC. By June 15, the sequence was already a story about who actually decides when a frontier AI model gets to exist in the world. Anthropic statement The Wall Street Journal: "Anthropic Dispatches Staff to D.C." CNBC: "Anthropic to meet with Trump administration"
Romero's frame for the whole episode is the word kayfable, defined in the column as a public performance maintained past the point of plausibility. The satirical register is his. Whether the sequence rises to that bar is a separate question. What is documented across the four beats is a pattern independent of the frame: Anthropic said no, did it anyway, and gave in only when the US government leaned on distribution. The Algorithmic Bridge: "Kayfable"
The pattern matters for a structural reason. The US government has no formal authority to retire an AI model. There is no Federal AI Commission with a rulebook. What it has is export controls, a tool built for arms and semiconductors, and the standing relationships of a national security apparatus. The June sequence shows what happens when that apparatus chooses to improvise: a directive goes out, a model comes down, and a frontier lab is sitting in the White House three days later. The mechanism is ad hoc, fast, and not written down anywhere a regulator would recognize. The Algorithmic Bridge: "Kayfable"
For Anthropic, the lesson of the six days is the lesson of Mythos and Fable 5 together: announcing a model as too dangerous to release, and then releasing a model from the same capability tier under guardrails, is a posture the US government is willing to override when the perceived risk reaches a threshold. The 72-hour turnaround on Fable 5 is the empirical fact every other frontier lab is now reading. The Algorithmic Bridge: "Kayfable"
For the US government, the lesson is that the export control tool works on a domestic AI lab in roughly the same way it works on a chipmaker. That is a capability, and capability, once demonstrated, gets used. The next release that touches biology, cybersecurity, or the perimeter of AI research will run on a different board than this one did. The Algorithmic Bridge: "Kayfable"
The open question, the one Romero's column ends on, is what Anthropic is actually optimizing for. The Mythos announcement, the Fable 5 release, the guardrail design, the 72-hour climbdown, and the White House meetings are six weeks of conduct, and a researcher, a customer, a regulator, and a competitor would each read them differently. The next lab in Anthropic's position will read them as data. What it learns from that data is the more durable story than the one that ended on June 12. The Algorithmic Bridge: "Kayfable"