Three thousand unpublished files. That's what was sitting in a publicly accessible data cache on Anthropic's website before anyone noticed.
A configuration error in Anthropic's content management system left digital assets set to public by default. Among the exposed files: draft blog posts describing an unreleased AI model called Claude Mythos, details about a planned CEO summit in Europe, internal images, PDFs, and even a document about an employee's parental leave. Anthropic pulled public access Thursday evening and blamed the exposure on "human error."
The model details are the headline. According to the leaked draft, Mythos belongs to a new tier Anthropic internally calls "Capybara" - described as "larger and more intelligent than our Opus models." The draft claims it achieves "dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity" compared to Claude Opus 4.6, which is currently Anthropic's most capable model.
What Anthropic Has Confirmed
Rather than deny the leak, Anthropic acknowledged the model's existence. A spokesperson told Fortune: "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. We consider this model a step change and the most capable we've built to date."
That's a careful statement, but it doesn't walk back the core claim. Anthropic is building something it considers significantly beyond Opus, and it's being cautious about release timing precisely because of how powerful it is.
The Cybersecurity Problem
The most alarming line from the leaked draft: 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."
That's Anthropic's own internal assessment, not an outside critic's hot take. The company that built the model is saying, in its own draft communications, that it could tip the balance between attackers and defenders in cybersecurity.
Wall Street took notice. CrowdStrike dropped 7% on Friday. Palo Alto Networks fell 6%. Zscaler lost 4.5%. Okta, SentinelOne, and Fortinet each shed about 3%. The logic is straightforward: if AI models get dramatically better at finding and exploiting software vulnerabilities, every cybersecurity company's job gets harder.
The Irony Is Hard to Miss
Anthropic is a company that has built its brand on AI safety. It publishes responsible scaling policies. It talks constantly about the risks of powerful AI systems. And it just accidentally exposed thousands of internal documents because someone didn't toggle a privacy setting in a CMS.
The leak itself wasn't a sophisticated attack - it was a checkbox left unchecked. That's the kind of basic operational security failure that Mythos, if the draft is accurate, would be very good at exploiting in other organizations' systems.
For anyone using Claude products today, nothing changes immediately. Mythos isn't released, and Anthropic says it's being "deliberate" about the rollout. But the confirmation that a model significantly more capable than Opus 4.6 is in testing is significant for anyone planning around AI capabilities in the coming months. And the cybersecurity implications deserve serious attention from anyone building software - not because of Mythos specifically, but because of the "wave" of similarly capable models Anthropic itself predicts is coming.