Anthropic’s next AI model is so dangerous it needs cyber-defender babysitters, yet the company leaked the whole plan through a mis-clicked CMS toggle.
The draft blog post, left in a publicly searchable data lake, names the model “Claude Capybara” and says it “gets dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity” than any previous Anthropic release.
A spokesperson confirmed the project is real:
“We’re developing a general purpose model with meaningful advances in reasoning, coding, and cybersecurity… the most capable we’ve built to date.”
According to the leaked document, Capybara 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 plans to release it first to “cyber defenders” so they can “improve the robustness of their codebases against the impending wave of AI-driven exploits.”
The leak also revealed an invite-only CEO retreat in the English countryside where Anthropic CEO Dario Amodei will demo unreleased Claude capabilities to European business leaders. After Fortune alerted the company, Anthropic blocked public access and blamed “human error” in its off-the-shelf content-management system, which defaults new uploads to public unless manually switched.
Close to 3,000 unpublished assets—including images, PDFs, and an employee’s parental-leave form—were exposed, according to University of Cambridge researcher Alexandre Pauwels.
The takeaway: Anthropic has trained a model it believes could supercharge hackers, but the first exploit was its own CMS.