Anthropic's Most Powerful AI Yet Was Leaked Before It Was Announced
Anthropic didn’t plan to tell you about Claude Mythos today. A human made a configuration error, and suddenly the world found out anyway.
On March 27, 2026, Fortune reported that Anthropic had accidentally left draft blog posts and nearly 3,000 unpublished assets in a publicly searchable data store. Cybersecurity researchers Alexandre Pauwels of the University of Cambridge and Roy Paz of LayerX Security stumbled upon the trove, reviewed the contents, and notified Fortune before Anthropic had a chance to announce anything officially. By the time Anthropic was informed and locked down the data store, the story was already out.
What those drafts revealed is significant — and not entirely comfortable reading.
Meet Claude Mythos (Also Known as Capybara)
The leaked draft blog post described a new model called Claude Mythos, which the document characterized as “by far the most powerful AI model we’ve ever developed.” In a statement to Fortune, Anthropic confirmed they’re testing the model, calling it “a step change” in performance and “the most capable we’ve built to date.”
The draft also introduced a new tier name: Capybara. If you’ve been following Anthropic’s naming conventions, here’s how the lineup is structured:
| Tier | Description |
|---|---|
| Haiku | Smallest, fastest, cheapest |
| Sonnet | Balanced performance and speed |
| Opus | Largest and most capable (until now) |
| Capybara | New tier — larger and more capable than Opus |
According to the leaked document, Capybara and Mythos appear to refer to the same underlying model — Mythos being the training name, Capybara being the product tier name. The document states: “Capybara gets dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity” compared to Claude Opus 4.6, the current best available model.
This isn’t an incremental upgrade. Anthropic is describing a structural change to their model lineup — a new category, not just a new version number.
The Cybersecurity Problem
Here’s where things get genuinely interesting, and a little unsettling.
Anthropic appears especially worried about what Mythos can do in the wrong hands. The leaked document described the model as “currently far ahead of any other AI model in cyber capabilities” and warned that it “presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders.”
In plain English: this model is so good at finding and exploiting security vulnerabilities that Anthropic is nervous about releasing it broadly.
It’s somewhat ironic, given that the company’s first public evidence of Mythos’s existence came via a security misconfiguration on their own end. Still, the concern appears genuine. Because of this risk profile, Anthropic says it’s releasing the model in early access exclusively to organizations focused on cyber defense — giving defenders a head start before the model’s capabilities become more widely available.
The company has prior experience navigating this kind of tension. Anthropic previously disrupted a Chinese state-sponsored campaign that had attempted to use Claude for malicious cyber operations, and has blocked exploitation attempts before. Mythos, by their own account, raises that risk profile significantly.
How the Leak Actually Happened
Anthropic attributed the incident to “human error” in the configuration of their content management system. A misconfiguration left draft materials — including what was clearly a structured, nearly publication-ready blog post — in a publicly accessible and searchable data store.
The company moved quickly once notified: the data store was locked down, and Anthropic issued a statement acknowledging the error and describing the leaked materials as “early drafts of content considered for publication.”
The same cache also included details of a planned invite-only CEO summit at an 18th-century English countryside manor, where Dario Amodei was set to meet European business leaders to discuss enterprise AI adoption. Anthropic called it “part of an ongoing series of events we’ve hosted over the past year.” A bit awkward to have that revealed alongside your flagship model, but here we are.
What This Means for the AI Race
The existence of a model tier above Opus is significant context for anyone watching the frontier AI space. OpenAI’s GPT-5 lineup has been expanding, Google’s Gemini 3 family was announced earlier this month, and the pressure on every major lab to demonstrate step-change improvements has never been higher.
Anthropic positioning Mythos/Capybara as genuinely different — not just better, but categorically more capable — signals the company believes it has something meaningful to show. Whether the benchmarks hold up to external scrutiny remains to be seen, since the leaked document is Anthropic’s own self-assessment.
The decision to initially restrict access to cybersecurity defenders is also worth watching. It suggests Anthropic is trying to build a moat in enterprise security before opening up consumer access — a commercially sensible move that also frames the release in terms of responsibility rather than competition.
What’s Next
Anthropic confirmed it’s working with a “small group of early access customers” to test the model. A broader release timeline hasn’t been announced — officially or accidentally. The company described itself as being “deliberate about how we release it,” which could mean weeks or months before the general public gets access.
Given how the announcement happened, expect Anthropic to now move toward a formal launch on their own terms. The draft blog post was clearly nearly ready. The only thing that changed is that someone else got to hit publish first.
Sources:
Share :
Stay Ahead in Tech
Join thousands of developers and tech enthusiasts. Get our top stories delivered safely to your inbox every week.
No spam. Unsubscribe at any time.
