Anthropic is preparing the most powerful AI in its history

Data Leak and Project Disclosure

The model description accidentally ended up in a public data cache and was analyzed by Fortune. A draft publication, stored in an unprotected repository, indicated that the new model is called Claude Mythos and is considered unprecedented in terms of cyber threats.

The same cache also revealed additional details—information about a closed CEO summit in Europe and materials related to promoting AI among corporate clients.

According to security researchers, approximately 3,000 resources related to the company’s blog, which had not been published before, became publicly accessible. After notification from Fortune, the company restricted access to the repository and admitted that the cause was a “human error” in the content management system.

Mythos and a New Tier of Models — Capybara

The documents also mention a new tier of models — Capybara, described as more powerful than the company’s current solutions. It appears that Capybara and Mythos belong to the same underlying system.

Currently, Anthropic’s model lineup is structured as follows: Opus as the most powerful models, Sonnet as a balance of speed and cost, and Haiku as the fastest and most affordable. According to the leak, the new Capybara tier will be above Opus—more performant but also significantly more expensive.

The documents also claim that the new model surpasses Claude Opus 4.6 in several tasks, including programming, academic reasoning, and cybersecurity. Training for Claude Mythos is reportedly already complete.

Cautious Launch

A general-purpose model is being developed with significant improvements in reasoning, code writing, and cybersecurity. At the same time, the company emphasizes that it is approaching the release with maximum caution.

The strategy includes a gradual rollout—starting with limited access for a small group of clients, followed by potential expansion. This approach is explained by the high operational cost of the model and its unpreparedness for mass use.

Significant Cybersecurity Risks

The new model carries substantial risks. Anthropic notes that the system already significantly outperforms other AI solutions in cybersecurity and may herald a wave of technologies capable of discovering and exploiting vulnerabilities faster than defensive systems.

In essence, this refers to a scenario where attackers could use such models to conduct large-scale attacks. This is precisely why the company plans to launch with a focus on protection—providing early access to organizations so they can prepare for a potential “wave of AI attacks.”

Rising Risks in the Industry

Similar concerns are observed among other market players, including OpenAI. As models advance, their ability to discover software vulnerabilities increases, creating a dual effect—technologies can be used both for defense and for attacks.

Anthropic has already encountered this issue—its models identified previously unknown vulnerabilities in real-world systems. The company also reported attempts by hacker groups, including those linked to state structures, to use Claude. In one instance, the attack affected approximately 30 organizations, including technology companies, financial institutions, and government agencies. Later, Anthropic blocked the accounts, conducted an investigation, and notified the affected parties.

Reason for the Leak

Experts link the incident to CMS features—uploaded files could become public by default if users did not change access settings. As a result, images, PDF files, and other materials ended up in the public domain.

Anthropic confirmed that the issue is related to an external content management tool and was caused by human error.