An unintentional content management system settings error has revealed the existence of Anthropic’s most powerful and advanced artificial intelligence model to date: an unannounced secret system named “Claude Mythos.” While the company asserts that this model represents an unprecedented qualitative leap in AI capabilities, it simultaneously fears its immense power could dangerously and unprecedentedly amplify cyberattacks.
Recent reports published by Fortune magazine indicated that precise descriptions of this secret model were found in a publicly accessible data storage repository containing blog post drafts and confidential internal documents. Cybersecurity researchers Roy Paz from LayerX Security and Alexander Bowles from the University of Cambridge independently managed to identify and review these sensitive materials. The leaks included nearly 3,000 unpublished files and resources associated with Anthropic’s blog. Following the magazine’s inquiry, the company quickly removed public access to the data store to limit the damage.
These leaks come at a critical time for the company as it seeks to strengthen its global presence. Reuters reported that the AI startup Anthropic is expanding its operations in Europe by opening new offices in major cities such as Paris and Munich, reflecting its major ambitions to dominate markets and deliver advanced technologies to a broader audience.
An entirely new class of artificial intelligence technologies
According to the leaked blog post draft, the “Claude Mythos” system introduces a new class of models called “Capybara,” described as larger, smarter, and more advanced than previous Opus models, which were until recently considered the company’s most powerful and intelligent creations.
The leaked document clearly points out that when comparing this new model to the Claude Opus 4.6 version—the company’s previous best—the Capybara model achieves significantly higher scores by a wide margin in software programming, academic reasoning, cybersecurity, and other complex technical domains.
In this regard, an Anthropic spokesperson confirmed that the company is currently developing a multi-purpose artificial intelligence model featuring significant and tangible advances in reasoning, programming, and cybersecurity. The spokesperson described the model as the most capable and efficient system the company has built to date. The company added that it is currently collaborating with a small, selected group of customers who have received early access to test the model and evaluate its performance in real-world work environments.
A warning bell and cybersecurity challenges
The leaked materials and documents paint a stark and complex picture of the dual-use capabilities of this advanced model. A draft from the company itself warned that the “Claude Capybara” model is currently miles ahead of any other available AI model in terms of cyber capabilities. The document explained that this system portends an upcoming wave of smart models possessing exceptional capabilities to exploit security vulnerabilities in ways that far exceed the efforts and capabilities of defenders and cybersecurity specialists.
To manage this grave risk and mitigate its fallout, the company has developed a structured strategic plan including the following points:
- Releasing the model initially exclusively to trusted cybersecurity organizations and companies.
- Granting defensive entities a proactive advantage and extra time to improve and enhance the robustness of their codebase.
- Preparing for the imminent wave of exploits and breaches driven by advanced artificial intelligence.
The leaked draft also noted that running this model requires very high costs due to the scale of its operations, and it has not yet reached the stage of being fully prepared for a public release.
A series of security concerns and escalating tensions
This serious leak comes amid mounting global tensions and concerns regarding AI technology security. Last February, Anthropic directed direct accusations against prominent Chinese AI laboratories—DeepSeek, Moonshot, and MiniMax—for creating over 24,000 fraudulent and fake accounts aimed at extracting and stealing AI model capabilities through a process known as distillation and knowledge copying.
In a related and separate context, the company revealed earlier that a group backed by Chinese state actors used an Anthropic software tool in an organized cyber campaign targeting nearly 30 global organizations, including major tech companies and sensitive government agencies, highlighting the growing security risks associated with these advanced tools.
Human error and future plans for technology leadership
The company attributed the exposure of this sensitive data to human error in its content management system settings. Because of this mistake, digital assets and files were set by default to public access unless manually restricted by technical administrators.
“Relying on default public access settings was a technical error that led to the early exposure of the company’s most important and advanced secrets.”
Alongside technical details, the cached storage also revealed secret plans to hold an exclusive executive retreat by invitation only. Scheduled to take place in an 18th-century historic English mansion, CEO Dario Amodei plans to showcase the capabilities of unannounced models to an elite group of business leaders and European officials in a move aimed at fostering partnerships and attracting strategic investments on the continent.
Frequently Asked Questions
What is the “Claude Mythos” model?
It is a highly secret and advanced artificial intelligence system developed by Anthropic, considered the most powerful and intelligent compared to the company’s previous models, and featuring an advanced new software class called “Capybara.”
How was the new AI model’s data leaked?
The leak occurred due to human error in the company’s content management system settings, where digital assets and internal drafts were left in default public access mode instead of being securely restricted.
What are the main concerns regarding this new model?
Primary concerns center around the model’s exceptional capabilities in cybersecurity, with the company fearing it could be used to launch complex cyberattacks exceeding the capabilities of current defense and protection systems.
How does the company plan to handle this model’s cybersecurity risks?
The company plans to release the model first to trusted cybersecurity organizations and defense entities to give them a proactive advantage in improving and enhancing their codebase rules and closing vulnerabilities before rolling the system out for general use.
