Anthropic Unveils Claude Mythos, an AI Model Considered Too Powerful

Claude Mythos: An AI Model Unveiled by Accident

An unexpected leak has brought to light Claude Mythos, an artificial intelligence model developed by Anthropic, which the company itself considers too powerful to be released to the general public. This revelation comes as the startup is contemplating an initial public offering (IPO). The model was discovered due to a configuration error that left nearly 3,000 files accessible in a public database. Among these documents, a draft describing the Claude Mythos model was spotted by two cybersecurity researchers, Alexandre Pauwels from the University of Cambridge and Roy Paz from LayerX Security. Furthermore, Bloomberg reported that Anthropic plans to go public by October 2026.

Capybara: A Model Beyond Opus

The leaked document presents a model called Capybara, which is actually another name for Claude Mythos. This model represents a new level in the Claude lineup, which already includes three versions: Haiku, Sonnet, and Opus, the latter being the most powerful to date. Capybara, or Mythos, is designed to surpass Opus in terms of size, cost, and capabilities. Performance tests indicate that Mythos excels in three key areas: programming, academic reasoning, and cybersecurity, with a particularly notable advantage in the latter. Anthropic has described the model as presenting unprecedented cyber risks.

A spokesperson for the company confirmed to Fortune that Mythos represents a significant leap in capabilities and is the most powerful model ever designed by Anthropic. However, it is currently being tested only by a select group of early-access clients, and no public launch date has been announced. The company is exercising caution, particularly due to the high operational costs associated with this model.

The leak also revealed an upcoming private summit in the UK, bringing together European CEOs and Dario Amodei, the co-founder of Anthropic. This underscores the company's intention to strengthen its presence in the European market.

A Tense IPO

The revelation of Claude Mythos comes at a strategic moment for Anthropic, which has engaged in discussions with investment banks in anticipation of an IPO planned for the fourth quarter. This timeline could place Anthropic in direct competition with OpenAI, which is also preparing for its public offering. Currently, Anthropic's valuation exceeds $300 billion, with an annualized revenue of $19 billion. However, the company's profitability remains uncertain, as the costs associated with training and operating the models consume a significant portion of the revenue.

This leak is particularly embarrassing for Anthropic, which is striving to build a reputation based on rigor and safety. The discovery of 3,000 unsecured files, described by a spokesperson as preliminary drafts, highlights flaws in the management of sensitive data. Additionally, the company is already embroiled in litigation with the Pentagon, complicating its position as it prepares to launch a model deemed too powerful for public release while also gearing up for its IPO.