Exclusive: Anthropic ‘Mythos’ AI mannequin representing ‘step change’ in energy revealed in knowledge leak

An Anthropic spokesperson stated the brand new mannequin represented “a step change” in AI efficiency and was “the most capable we’ve built to date.” The firm stated the mannequin is at present being trialed by “early access customers.”

Descriptions of the mannequin have been inadvertently saved in a publicly-accessible knowledge cache and have been reviewed by Fortune.

A draft weblog submit that was accessible in an unsecured and publicly-searchable knowledge retailer previous to Thursday night stated the brand new mannequin known as “Claude Mythos” and that the corporate believes it poses unprecedented cybersecurity dangers.

The identical cache of unsecured, publicly discoverable paperwork revealed particulars of a deliberate, invite-only CEO summit in Europe that’s a part of the corporate’s drive to promote its AI fashions to massive company prospects. 

The AI lab left the fabric, together with what gave the impression to be a draft weblog submit asserting a brand new mannequin, in an unsecured, public knowledge lake, based on paperwork individually positioned and reviewed by Roy Paz, a senior AI safety researcher at LayerX Security, a pc and community safety firm, and Alexandre Pauwels, a cybersecurity researcher on the University of Cambridge. 

In complete, there gave the impression to be shut to three,000 property linked to Anthropic’s weblog that had not been revealed beforehand on the corporate’s information or analysis websites that have been nonetheless publicly-accessible on this knowledge cache, based on Pauwels, who Fortune requested to evaluate and overview the fabric.

After being knowledgeable of the information leak by Fortune on Thursday, Anthropic eliminated the general public’s capability to go looking the information retailer and retrieve paperwork from it.

In an announcement offered to Fortune, Anthropic acknowledged {that a} “human error” within the configuration of its content material administration system led the draft weblog submit to being accessible. It described the unpublished materials that was left in an unsecured and publicly-searchable knowledge retailer as “early drafts of content considered for publication.”

As effectively as referring to Mythos, the draft weblog submit additionally mentioned a brand new tier of AI fashions that it says will likely be referred to as “Capybara”. In the doc, Anthropic says: “’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.” Capybara and Mythos seem to confer with the identical underlying mannequin.

Currently, Anthropic markets every of its fashions in three completely different sizes: the most important and most succesful mannequin variations are branded Opus, whereas a barely sooner and cheaper, however much less succesful, variations are branded Sonnet, and the smallest, most cost-effective, and quickest are referred to as Haiku. However, within the weblog submit, Anthropic describes Capybara as a brand new tier of mannequin that’s even bigger and extra succesful than Opus, but additionally dearer.

“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 corporate stated within the weblog.

The doc additionally stated the corporate had accomplished coaching “Claude Mythos,” which the draft weblog submit described as “by far the most powerful AI model we’ve ever developed.”

In response to questions concerning the draft weblog submit, the corporate acknowledged coaching and testing a brand new mannequin. “We’re developing a general purpose model with meaningful advances in reasoning, coding, and cybersecurity,” an Anthropic spokesperson stated. “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 doc Fortune and the cybersecurity specialists reviewed consists of structured knowledge for a webpage, full with headings and a publication date, suggesting it kinds a part of a deliberate product launch. It outlines a cautious rollout technique for the mannequin, starting with a small group of early-access customers. The draft weblog notes that the mannequin is pricey to run and never but prepared for basic launch.

Significant new cybersecurity dangers

The new AI mannequin poses vital cybersecurity dangers, based on the leaked doc. 

“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 doc stated.

Anthropic seems to be particularly fearful concerning the mannequin’s cybersecurity implications, noting that the system is “currently far ahead of any other AI model in cyber capabilities” and “it presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders.” In different phrases, Anthropic is anxious that hackers may use the mannequin to run large-scale cyberattacks.

The firm stated within the draft weblog that due to this danger, its plan for the mannequin’s launch would give attention to cyber defenders: “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 newest era of frontier fashions from each Anthropic and OpenAI have crossed a threshold that the businesses say poses new cybersecurity dangers. In February, when OpenAI launched GPT-5.3-Codex, the corporate said it was the primary mannequin it had categorized as “high capability” for cybersecurity-related duties beneath its Preparedness Framework—and the primary it had immediately educated to determine software program vulnerabilities. 

Anthropic, in the meantime, navigated similar risks with its Opus 4.6, launched the identical week. The mannequin demonstrated a capability to floor beforehand unknown vulnerabilities in manufacturing codebases, a functionality that the corporate acknowledged was dual-use, that means that it may each assist hackers in addition to assist cybersecurity defenders discover and shut vulnerabilities in code.

The firm has additionally reported that hacking teams, together with these linked to the Chinese authorities, have tried to use Claude in real-world cyberattacks. In one documented case, Anthropic found {that a} Chinese state-sponsored group had already been working a coordinated marketing campaign utilizing Claude Code to infiltrate roughly 30 organizations—together with tech corporations, monetary establishments, and authorities businesses—earlier than the corporate detected it. Over the next ten days, Anthropic investigated the total scope of the operation, banned the accounts concerned, and notified affected organizations.

An unique govt retreat

The leak of not-yet-public data seems to stem from an error on the a part of customers of the corporate’s content material administration system (CMS), which is the software program used to publish the corporate’s public weblog, based on cybersecurity professionals. 

Digital property created utilizing the content material administration system are set to public by default and usually assigned a publicly accessible URL when uploaded—until the consumer explicitly modifications a setting in order that these property are stored personal. As a outcome, a big cache of pictures, PDF recordsdata, and audio recordsdata appear to have been revealed erroneously to an unsecured and publicly-accessible URL through the off-the-shelf content material administration system.

Anthropic acknowledged in an announcement to Fortune that “an issue with one of our external CMS tools led to draft content being accessible.” It attributed this concern to “human error.” 

Many of the paperwork gave the impression to be discarded or unused property for previous weblog posts like pictures, banners, and logos. However, a number of gave the impression to be what have been meant to be personal or inside paperwork. For instance, one asset has a title that described an worker’s “parental leave.” 

The paperwork additionally included a PDF containing details about an upcoming, invite-only retreat for the CEOs of European corporations being held within the U.Okay., and which Anthropic CEO Dario Amodei will attend. Names of the opposite attendees aren’t listed, however are described as Europe’s most influential enterprise leaders.

The two-day retreat is described as an “intimate gathering” to have interaction in “thoughtful conversation” at an 18th-century manor-turned-hotel-and-spa within the English countryside. The doc says that attendees will hear from lawmakers and policymakers about how companies are adopting AI and expertise unreleased Claude capabilities.

An Anthropic spokesperson informed Fortune the occasion “is part of an ongoing series of events we’ve hosted over the past year. We look forward to hosting European business leaders to discuss the future of AI.”