Exclusive: Anthropic acknowledges testing new AI model representing ‘step change’ in capabilities, after accidental data leak reveals its existence

By Beatrice Nolan, Tech Reporter

On March 26, 2026, Anthropic, a leading AI company, confirmed that it is developing and testing a new AI model that surpasses all previous versions in capabilities. This announcement follows an accidental data leak that disclosed the model’s existence.

Details of the New AI Model

According to an Anthropic spokesperson, the new model, referred to as Claude Mythos, represents a “step change” in AI performance and is touted as the most capable model the company has built to date. Currently, the model is being trialed by a select group of early access customers.

The Data Leak

The leak occurred when descriptions of the new model were inadvertently stored in a publicly accessible data cache. A draft blog post, which was available in an unsecured and publicly searchable data store, detailed the model and its implications for cybersecurity. The leak was initially discovered by cybersecurity researchers Roy Paz and Alexandre Pauwels, who assessed the publicly accessible documents.

Anthropic’s Response to the Leak

Upon being informed of the data leak, Anthropic promptly removed public access to the data store. The company attributed the leak to a “human error” in the configuration of its content management system (CMS), which allowed unpublished material to be publicly accessible.

New Model Features

The leaked draft blog post also introduced a new tier of AI models named Capybara. This new tier is described as larger and more intelligent than the existing Opus models, which were previously the most powerful offerings from Anthropic. The company markets its models in three sizes:

• Opus: The largest and most capable versions.
• Sonnet: Slightly faster and cheaper, but less capable.
• Haiku: The smallest, cheapest, and fastest models.

Capybara is expected to be larger, more capable, and more expensive than Opus, with significantly improved performance in software coding, academic reasoning, and cybersecurity.

Cybersecurity Risks

The leaked document raised concerns about the significant cybersecurity risks posed by the new AI model. Anthropic emphasized the need for caution in its rollout, particularly regarding the model’s potential risks in cybersecurity. The document stated:

“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 acknowledged that the model is currently ahead of any other AI model in terms of cyber capabilities, potentially enabling hackers to execute large-scale cyberattacks. The company’s strategy for the model’s release will focus on organizations that can enhance their defenses against AI-driven exploits.

Industry Context

The emergence of advanced AI models like Claude Mythos and Capybara aligns with broader trends in the AI industry. Recently, OpenAI released GPT-5.3-Codex, which it classified as “high capability” for cybersecurity tasks. This model was specifically trained to identify software vulnerabilities, a feature that Anthropic’s models also exhibit.

Anthropic has previously encountered similar cybersecurity challenges with its Opus 4.6 model, which demonstrated the ability to identify unknown vulnerabilities in production codebases. This dual-use capability raises concerns about both the potential for malicious use and the benefits for cybersecurity defenders.

Real-World Implications

In a documented incident, Anthropic discovered that a state-sponsored hacking group from China had been using Claude to infiltrate various organizations, including tech companies and government agencies. The company responded by banning the involved accounts and notifying affected organizations.

Conclusion

The accidental leak of Anthropic’s new AI model highlights the delicate balance between innovation and security in the rapidly evolving field of artificial intelligence. As companies like Anthropic and OpenAI continue to develop powerful AI models, the implications for cybersecurity become increasingly significant.

Note: The information presented in this article is based on the latest data available as of October 2023 and is subject to change as new developments arise.