Sometimes, it’s not an official announcement but a small mistake that reveals the biggest developments.

That’s exactly what happened with Anthropic, after a data leak quietly exposed details about a powerful new AI model the company has been working on.

The model, referred to as Claude Mythos, surfaced after internal files were accidentally left accessible in a public cache. The information was first noticed and reviewed by Fortune, offering a rare look at a project that hasn’t been formally announced yet.

Anthropic later confirmed that the model is real. According to the company, it is building a general-purpose AI system with stronger abilities in reasoning, coding, and cybersecurity.

From the details that emerged, Claude Mythos appears to go well beyond Anthropic’s current models. Internally, it is said to be part of a new “Capybara” tier, a level above its existing Opus systems, both in capability and cost.

The leak itself was traced back to a simple error. The company said some early draft content, including what looked like a blog post, was mistakenly made public due to a configuration issue in its content system. Access was later restricted once the problem was identified.

What stood out in the leaked material was not just the model’s power, but the concerns around it. Anthropic reportedly noted that its cyber capabilities are already ahead of other systems, raising the possibility that such models could be used to find and exploit software weaknesses faster than defenders can respond.

That risk seems to be shaping how the company is rolling it out. Early testing is believed to be limited, with priority given to organisations working in cyber defence, likely to better understand and prepare for possible misuse.

For now, many details about Claude Mythos remain unclear, and Anthropic has not made any official public launch announcement.

At the same time, the company is also said to be preparing for a major step on the business side. According to The Information, Anthropic could explore going public as early as late 2026, with expectations of raising over $60 billion—potentially one of the largest IPOs in recent years.

Also Read: Anthropic Secures Injunction Against Pentagon in Landmark AI Safety Showdown