A misconfiguration in Anthropic’s content management system exposed nearly 3,000 unpublished assets to public search on March 26, 2026. Among them was a draft blog post describing a new AI model the company had not yet announced. The model is called Claude Mythos. Anthropic confirmed its existence to Fortune after being notified of the leak, describing it as “a step change” in AI performance and “the most capable we’ve built to date.”

The company attributed the incident to human error. Assets published through its CMS are set to public by default unless a user explicitly changes that setting. The exposed cache was discovered by Roy Paz, a senior AI security researcher at LayerX Security, and Alexandre Pauwels, a cybersecurity researcher at the University of Cambridge. Anthropic removed public access after Fortune contacted the company.

What the documents actually say

The leaked draft refers to the model under two names. “Mythos” appears in the first version of the post. A second version swaps it for “Capybara” throughout — though the subtitle of that version still reads “We have finished training a new AI model: Claude Mythos,” suggesting the company was deciding between names rather than describing two separate systems.

Capybara, according to the draft, is described as “a new name for a new tier of model: larger and more intelligent than our Opus models — which were, until now, our most powerful.” Anthropic currently structures its lineup in three tiers: Haiku, Sonnet, and Opus. Mythos would sit above all three.

The performance claims are specific. “Compared to our previous best model, Claude Opus 4.6, Capybara gets dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity, among others,” the draft states. An Anthropic spokesperson confirmed to Fortune that the model offers “meaningful advances in reasoning, coding, and cybersecurity.”

The cybersecurity problem Anthropic flagged itself

The more unusual element of the leak is that Anthropic’s own draft document warns about the model’s risks. According to the materials reviewed by Fortune, Mythos is “currently far ahead of any other AI model in cyber capabilities” and “presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders.”

The company says it plans a deliberately slower release than with previous models, starting with a small group of early-access customers evaluating cybersecurity applications before any broader API rollout. The draft also acknowledges that the model is “very expensive for us to serve, and will be very expensive for our customers to use,” with Anthropic working to improve efficiency before general availability.

The irony of a company building a model it describes as posing unprecedented cybersecurity risks — and then exposing that model’s existence through a basic CMS configuration error — has not been lost on the industry.

What it means for the competitive landscape

March 2026 has already seen GPT-5.4 and Gemini 3.1 Pro trade benchmark leads across coding and reasoning tasks. A new Anthropic tier above Opus, if the leaked performance claims hold under independent testing, would reset the frontier again. The timing also fits a pattern: Anthropic has been accelerating releases since in February, and the existence of an active early-access program suggests Mythos is closer to release than the leak makes it appear.

Whether the model ships as Mythos, Capybara, or under a third name remains unclear. What is clear is that Anthropic is testing something it considers materially more capable than anything currently available — and is being careful about how it reaches the public.