Security researchers Roy Paz of LayerX Security and Alexandre Pauwels of the University of Cambridge stumbled upon a publicly accessible Anthropic data repository. Inside? Nearly 3,000 unpublished files, including a draft blog post describing a model whose existence was not yet known to the world. It is called Claude Mythos, and its leak has unleashed quite a storm.
Anthropic quickly acknowledged the error. Its content management system made newly uploaded files public unless the user manually switched them to private. Simply human error. But the consequences of that error are anything but trivial.
What will Mythos be like?
According to the leaked document reviewed by Fortune, Anthropic describes Mythos as "the most powerful AI model it has ever developed". The name Capybara is also used internally, with both names apparently referring to the same model. Mythos is expected to introduce an entirely new tier above the existing Opus line. Until now, the hierarchy consisted of Haiku, Sonnet, and Opus, with Opus being the most powerful. Mythos will become the new top of the pyramid. It is a larger, more expensive, and more powerful model than Opus 4.6, aimed primarily at enterprise customers with complex requirements.
What exactly does it do better? According to the leaked comparison, its results in programming, academic reasoning, and cybersecurity improved significantly. On SWE-bench Verified, a benchmark that tests a model's ability to fix real bugs in production software projects, Opus 4.6 achieved a score of around 72 to 73 percent. Mythos was reportedly scoring in the mid-to-high 80s.
Powerful in cybersecurity, which is a problem
And now for the part that is keeping experts awake at night. The leaked document states directly that Mythos is "currently far ahead of any other AI model in cyber capabilities" and heralds a wave of systems that will discover vulnerabilities far faster than defenders can respond.
Anthropic itself describes this in the document as an unprecedented risk. The model can identify security flaws in software with a level of accuracy that was previously impossible. For security teams, penetration testers, and researchers, it is an extremely effective tool. But in the wrong hands, it becomes a tool for large-scale automated attacks.
Incidentally, Anthropic has direct experience with this issue. The company previously documented a case in which a Chinese state-sponsored group used Claude to infiltrate approximately 30 organizations, including technology companies, financial institutions, and government agencies. Anthropic detected the campaign, blocked the accounts, and informed the affected organizations. This experience is clearly shaping how it is approaching the release of Mythos.
That is why the company has chosen a cautious path: for now, access is limited to a small group of early-access customers, focused specifically on cyber defenders so that they can get a head start before the wave of AI attacks arrives.
How the data leak occurred
The technical failure was surprisingly simple. The content management system Anthropic uses for its blog automatically made newly uploaded files public with an accessible URL. If a user forgot to flip the switch, the file was available to anyone who knew where to look.
Data security experts point out that this could easily have been prevented. Tools such as DSPM (data security posture management) would have automatically detected the misconfigured storage and flagged the leaked files as sensitive. DLP (data loss prevention) systems, meanwhile, would have blocked public addresses from being assigned to files containing internal code names such as "Mythos" or "Capybara". The files simply would not have left the secure environment.
Anthropic blocked access to the data repository immediately after being alerted by Fortune. But Pauwels and Fortune had enough time to examine the documents, and the news spread around the world.
What we still do not know about Mythos
The leaked figures are enticing, but they need to be viewed with a clear head. Benchmarks published before a model's release are selected by the manufacturer itself and presented in the most favorable light. The final version of the model may look different when released. Moreover, even the best score does not mean that the model will work in practice exactly as the tests suggest.
We still do not know the exact release date. We do not know the price or technical specifications such as the context window length. Anthropic has only confirmed that it is testing the model with selected customers and considers it "a step change in quality and the most capable model it has ever developed".
There is also speculation about the existence of a secondary model in the same line, tentatively called Claude Capiara, which could serve a role similar to that of the smaller variants in the Opus line. However, Anthropic has not yet confirmed this.
