AI-pocalypse. The main thing about the scandal with a secret AI model that can bring down anything

At the end of March 2026, after several leaks, the world learned of the existence of a secret AI model from Anthropic that could collapse global financial markets and destroy an entire industry. Claude Mythos turned out to be able to find holes in programs in a matter of minutes that neither IT security specialists nor hackers had noticed for years. Shares of cybersecurity companies immediately collapsed, despite the fact that the company promised never to release Mythos to the public. The market panic was so severe that Congressional hearings were requested and the White House held an emergency meeting on the issue. Should we expect an AI-pocalypse in the near future or will Mythos turn out to be just a myth? - Yana Zarechnaya was looking into it.

Myth or capybara

It was an ordinary March evening, which did not foretell great journalistic successes. Fortune magazine writer Beatrice Nolan routinely searched URLs in the blogging infrastructure of Anthropic, the same company that developed the chatbot Claude, in search of a new topic. This is a rather tedious matter: on the company’s website, each uploaded document is assigned a public web address by default - Google does the same with your documents, unless the author has prohibited this; drafts and some internal materials can be read by anyone if they select a link. By the way, remember this when you share documents with your friends.

Journalists covering the affairs of tech corporations use this method of “intelligence” regularly. It even received the name directory enumeration (literal translation - enumeration of directories). But it rarely brings results - corporations know how to keep secrets.

Not in this case.

Searching through the links, Nolan came across a repository of about three thousand unpublished internal Anthropic materials: draft blog posts, presentations, marketing documents, details of a closed summit for European CEOs of large companies. Among this digital Eldorado, she found a real treasure - a draft blog post about the creation of a new Anthropic model.

Under the headline “Introducing Capybara,” it said, “We estimate that this model heralds a coming wave of AI capable of exploiting vulnerabilities at a speed that defenders will not be able to match.” The developers called the new model Capybara, and planned to sell it under the name Claude Mythos.

Beatrice Nolan's article was published on the evening of March 26 - the next day the whole world knew about the sensation.

Subscribe to read in full

Subscribe to Redefine.Media to read Republic

Subscribe [You can pay with a Russian or foreign card. The subscription is renewed automatically. You can unsubscribe at any time.]

Where does the subscribers' money go?

Most Republic content is available through a paid subscription. We think this is a good way to finance media. After all, as you know, if you didn’t pay to read it, then someone else paid for you to read it. In our case, everything is fair: your money pays salaries and fees to journalists, and they write about topics that are important and interesting to you.

A key feature of our subscription: Your money is split across Republic magazines based on how you read them. If you read materials from one magazine, then your money will go only to him, and will not go to others. That is, you finance only what interests you.

Republic uses a Redefine.Media subscription. To register, we will redirect you to the Redefine.Media website, where you will need to register and pay for a subscription. Logging into the Redefine.Media website will allow you to read Republic materials from the same device.

A yearly subscription is more profitable than a monthly subscription. And if you want to unsubscribe, you can always do this in your personal account.