Anthropic Leaks AI That Could Threaten the Internet Due to Configuration Error

The Anthropic company accidentally exposed nearly 3,000 internal documents online due to a configuration error in their content management system, a fact that came to light on the evening of March 26, 2026. Among the leaked documents was the official draft of the Claude Mythos, a new model that Anthropic describes as “By far, the most powerful AI model we've ever developed.”

According to a report by Fortune, which reviewed the materials and notified the company, the draft mentions an internal tier called Capybara and highlights significant advancements in reasoning, coding, and especially cybersecurity. The internal document warns that the model is “far ahead of any other AI in cyber capabilities” and could pave the way for a wave of automated attacks that outstrip current defenses.

Practically speaking, the leaked benchmarks show Claude Mythos far ahead of Claude Opus 4.6 in these areas. Anthropic confirmed the leak as a “human error” and said they restricted access as soon as they were notified. The company has already begun testing the model with a small group of early access clients but emphasizes caution due to the security risks mentioned in the document.

The incident is somewhat ironic, as Anthropic positions itself as one of the most security-focused and AI-aligned companies, yet a simple configuration error exposed roadmaps, images, and entire drafts. Security researchers from LayerX and Cambridge University were the first to identify the public endpoint.

The company plans to continue controlled testing of Claude Mythos in the coming weeks and to define a commercial launch schedule, with extra emphasis on the cyber risk mitigation protocols that the leak has already highlighted.