Anthropic has announced the development and ongoing testing of a groundbreaking AI model, which surpasses all previous iterations in terms of capability. This confirmation follows a notable data leak that disclosed the model's existence.
A spokesperson from Anthropic described this new model as "a step change" in AI performance, asserting it to be "the most capable we’ve built to date." Currently, the model is being evaluated by "early access customers."
The details pertaining to the model were inadvertently stored in a publicly accessible data cache, which was examined by Fortune.
Prior to Thursday evening, an unprotected draft blog post detailing the new model was found in an unsecured and publicly-searchable data repository. The model, referred to as "Claude Mythos," has been identified by the company as potentially presenting unprecedented cybersecurity challenges.
Additionally, the same data cache disclosed information about an exclusive, invite-only CEO summit scheduled in Europe. This event is part of Anthropic's strategic approach to market its AI models to major corporate clients.
The vulnerability in the AI lab's data management was highlighted by Roy Paz, a senior AI security researcher at LayerX Security, and Alexandre Pauwels, a cybersecurity researcher affiliated with the University of Cambridge. They identified the unsecured documents, which included a draft blog post about the new model.
Pauwels outlined that nearly 3,000 assets linked to Anthropic’s blog, which had not been made public via the company’s official news or research outlets, were found in this data cache.
Upon notification from Fortune regarding the data leak, Anthropic promptly restricted public access to the data store to prevent further document retrieval.
In a statement to Fortune, Anthropic attributed the leak to "human error" in the setup of its content management system, describing the accessible unpublished material as "early drafts of content considered for publication."
Aside from mentioning Mythos, the draft blog post also introduced a new tier of AI models named "Capybara." According to the document, Anthropic stated: "‘Capybara’ is a new name for a new tier of model: larger and more intelligent than our Opus models—which were, until now, our most powerful." It seems that Capybara and Mythos refer to the same underlying architecture.
Currently, Anthropic categorizes its models into three sizes: the most advanced versions are branded as Opus; intermediate models offering slightly faster performance at a lower cost are called Sonnet; and the smallest and most economical models are designated as Haiku. However, the blog post suggests that Capybara will represent a new model tier that is not only larger and more capable than Opus but also more costly.
The company claims that Capybara significantly outperforms its predecessor, Claude Opus 4.6, scoring substantially higher in tests related to software development, academic reasoning, and cybersecurity, among other fields.
The leaked document also reveals that Anthropic has finalized the training of "Claude Mythos," characterizing it as "by far the most powerful AI model we’ve ever developed."
In response to inquiries about the draft blog post, Anthropic confirmed the ongoing training and evaluation of an advanced model. A spokesperson noted, "We’re developing a general purpose model with meaningful advances in reasoning, coding, and cybersecurity. Given the strength of its capabilities, we’re being deliberate about how we release it. As is standard practice across the industry, we’re working with a small group of early access customers to test the model. We consider this model a step change and the most capable we’ve built to date."
The document reviewed by Fortune and cybersecurity experts showcased structured web data, complete with headings and a publication date, indicating that it was part of a planned product launch. It emphasized a cautious rollout approach for the model, initiating with a select group of early-access users. Furthermore, the document highlighted that the operational costs of the model are significant and that it is not yet prepared for a broader release.
Serious cybersecurity concerns
According to the leaked document, the new AI model poses considerable cybersecurity threats.
"In preparing to release Claude Capybara, we want to act with extra caution and understand the risks it poses—even beyond what we learn in our own testing. In particular, we want to understand the model’s potential near-term risks in the realm of cybersecurity—and share the results to help cyber defenders prepare," the document stated.
Anthropic has expressed particular concern over the new model's cybersecurity implications, noting that it is "currently far ahead of any other AI model in cyber capabilities" and may herald a wave of models capable of exploiting vulnerabilities at a pace that outstrips defenders' countermeasures. The company worries that malicious actors might leverage this model to conduct large-scale cyberattacks.
The draft blog further elucidated that due to these risks, the approach for the model’s rollout would be primarily directed toward cyber defense organizations: "We’re releasing it in early access to organizations, giving them a head start in improving the robustness of their codebases against the impending wave of AI-driven exploits."
The advancements in AI models, both from Anthropic and OpenAI, have crossed thresholds that the companies assert create new cybersecurity risks. In February, OpenAI's launch of GPT-5.3-Codex was categorized as the first model classed as "high capability" for cybersecurity-related tasks within its Preparedness Framework, highlighting its direct training focused on identifying software vulnerabilities.
Simultaneously, Anthropic has navigated similar cybersecurity concerns with its Opus 4.6 model.
Share this story