U.S. artificial intelligence (AI) company Anthropic has reportedly developed a next-generation model for its large language model (LLM) “Claude” that surpasses its existing versions, and has begun testing it with early customers.

According to U.S. business outlet Fortune and others on the 29th, an internal document was accidentally exposed on Anthropic’s online repository. During the operation of an external content management system (CMS), a staff error resulted in the relevant file being uploaded to an internet address (URL) with no security settings. Anthropic explained this as “human error.”

According to the leaked blog draft, the new model is named “Claude Mythos” and is classified under a new tier called “Capybara,” which exceeds the performance of its current top model, Opus. In the draft, Anthropic stated, “Compared with the previous top model Claude Opus 4.6, it recorded significantly higher scores on various tests including software coding, academic reasoning, and cybersecurity.”

After the leak became known, an Anthropic spokesperson also acknowledged that this model represents a “step-change leap” in AI performance and is “the best model we have built so far.” Anthropic believes that the model’s capabilities in both cyberattack and cyber defense far outstrip those of existing AI, raising the risk that hackers could exploit it, and therefore plans to initially make it available primarily to cybersecurity companies and similar entities.

Popular News