The State Called Anthropic
This week, a San Francisco company received twenty-five billion dollars from Amazon, obtained the exclusive use of the United States National Security Agency, locked out eight European cybersecurity agencies, overwhelmed open-source maintainers with vulnerability reports they had not requested, and began collecting, by silent decree, the documents of three hundred thousand companies that use Atlassian. Last week it was a company. This week it is something else.
Amazon paid its protection money in advance. On Monday, April 20, Amazon announced up to twenty-five billion dollars more in Anthropic, on top of the eight billion already committed. Five billion arrives immediately; the other twenty is tied to "commercial milestones," a phrase that on Wall Street means everything and nothing. In exchange, Anthropic has committed to spending one hundred billion dollars on AWS over the next ten years, and to running Claude on more than one million Trainium2 chips that Amazon designed. "Our users tell us Claude is increasingly essential to how they work," said Dario Amodei. Essential things, apparently, now have a price tag.
The Department of War considers Anthropic a "supply chain risk." The National Security Agency, for its part, is already using it. Axios reported on April 19 that the NSA is running Mythos, Anthropic's internal cyber model, at the same time that its parent department has flagged the company as dangerous. The origin of the quarrel: the Department wanted Claude to serve "all lawful purposes" — a category that, in the plain-English sense, includes mass domestic surveillance and autonomous weapons. Anthropic refused. Two days before the Axios report, Amodei met in Washington with Susie Wiles, the White House Chief of Staff, and with Scott Bessent, the Treasury Secretary. No press release followed. What followed, it seems, was an understanding.
Meanwhile, in another corner of the map, open-source software is being conscripted, without its consent, into Anthropic's quality-control department. For weeks Mythos has flooded the maintainers of volunteer projects with vulnerability reports no one asked for. Daniel Stenberg, the Swedish engineer who maintains cURL — the piece of software that is most likely delivering this web page to the reader's browser — received one hundred and eighty-one notifications last year, spread across himself and his seven-person volunteer team. In the first three months of 2026, cURL has found and fixed more vulnerabilities than in either of the two previous full years. Fewer than one percent of discovered flaws have been patched. Mythos reported, among other things, a twenty-seven-year-old bug in OpenBSD. A twenty-seven-year-old bug is reassuring. The rest are something else.
Eight European agencies do not have access to the model. The United Kingdom does. Politico reported that Anthropic has granted Mythos Preview to about forty organizations, almost all in allied territory. The European Commission is not among them. Germany — the continent's largest economy — is "in talks," according to the sources. The British AI Security Institute has already evaluated it: Mythos scored seventy-three percent on expert-level capture-the-flag challenges, and solved, end to end, three out of ten simulations of a thirty-two-step corporate network attack. Minister Kanishka Narayan confirmed his government "has already taken action" based on the findings. What action, he did not say. That part, apparently, is not for every country.
Starting August seventeenth, Atlassian will begin reading the documents of three hundred thousand companies. The Australian firm announced that on that date it will begin collecting, by default, the metadata and in-app contents of Jira and Confluence to train its AI products. Free and Standard plans are not permitted to disable metadata collection. Retention will run up to seven years. Separately, and adjacent, a Reddit account created three days earlier claimed that Anthropic was in advanced talks to buy Atlassian at one hundred and fifty dollars per share. The account had a single post. Hacker News discussed it at length. No credible outlet has confirmed anything. The interesting part is not the rumor. The interesting part is how many people found it plausible.