Anthropic leak reveals cybersecurity danger and potential of new model

Mythos is a font of information.

That’s the concern for cybersecurity pros reviewing information about Anthropic’s new AI model, Mythos, part of a new tier of models named Capybara. According to leaked product information meant for early access customers, the new model could be used by threat actors to implement more powerful attacks.

Leaky faucets. Fortune first reported on the leak on March 26. Almost 3,000 documents on the model, including images and PDFs, were accidentally made available on Anthropic’s blog CMS and accessible to the public without needing a login (they’ve since been taken down). The content was meant to give customers a preview of the new model’s capabilities.

Mythos allegedly represents a “step change” in capabilities and the “most powerful” model Anthropic says it has ever built, the company told Fortune in a statement after the leak.

Danger, danger. Anthropic claims that attackers can leverage the new tier’s power to exploit cybersecurity vulnerabilities.

“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,” Anthropic’s documents claim. “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.”

Cybersecurity companies are taking notice, and so are investors. The market saw declines in cybersecurity-related stocks, with Zscaler in particular taking a heavy hit Monday morning. The company is down around 40% compared to the beginning of the year.

Future, modeled. Mythos is likely just the beginning. AI executives told Axios CEO Jim VandeHei that OpenAI and other big tech companies are expected to debut their own high-capability versions of AI models that could be deployed by attackers. OpenAI’s upcoming model, codenamed Spud, is said to rival Mythos.

Lian Jye Su, an Omdia analyst, told AI Business that Anthropic is invested in cybersecurity companies using models like Mythos to stay ahead of threats. It is “already a known fact that most cybersecurity companies will be using models from this foundation model provider,” he said, making it an important part of Anthropic’s business.

“It points more toward a future where AI now becomes a lot more embedded in various applications,” Su added.