Artificial intelligence company Anthropic released a preview version of its new "frontier model," Mythos, on Tuesday. The model is called "the most powerful to date" and will be first given to a small group of partners through a security initiative called "Project Glasswing," focusing on cybersecurity tasks.

According to the company's introduction, Mythos is a general-purpose large model in the Claude system, with outstanding programming and reasoning capabilities. Although not specifically trained for cybersecurity, the current preview version is mainly used to scan code vulnerabilities in first-party and open-source software. Anthropic stated that Mythos has discovered "thousands of zero-day vulnerabilities" in recent tests, many of which are highly dangerous, and some can be traced back to old code from one to twenty years ago. This discovery demonstrates the huge potential of large models in uncovering long-ignored "historical vulnerabilities."

image.png

However, with this technological advancement, concerns about its security have also emerged in the industry, especially the potential threats if these tools are misused. To control potential risks, Anthropic clearly stated that this preview version is not open to the public and is only available for use by signed partners in specific security tasks.

Currently, organizations participating in the Mythos trial include dozens of institutions such as Amazon, Apple, Broadcom, Cisco, CrowdStrike, Linux Foundation, Microsoft, and Palo Alto Networks. These organizations will jointly use the model for "defensive security work" and share experiences after the project ends to help the entire technology industry improve security levels.

In addition, Anthropic has been engaged in "ongoing discussions" with U.S. federal officials regarding the potential applications of Mythos in national critical systems. This discussion coincides with Anthropic's current legal dispute with the U.S. Department of Defense, which listed the company as a "supply chain risk entity" because the company refused to allow its technology to be used for autonomous strikes or monitoring of U.S. citizens.

Notably, the exposure of Mythos originated from a data security incident. Last month, an unpublished blog draft from Anthropic was mistakenly stored in a publicly accessible data lake, leading to rapid discovery and dissemination by security researchers. The draft mentioned that Mythos is a larger and smarter model than the previous Opus series and warned that if the model is misused, it could discover and exploit vulnerabilities at a speed faster than defenders can fix them.

Key Points:

🌟 Mythos is the latest AI model launched by Anthropic, currently available for trial use by a limited number of security partners.  

🔍 The model has uncovered thousands of zero-day vulnerabilities during testing, demonstrating its potential in uncovering historical vulnerabilities.  

⚠️ The company is discussing its application in national critical systems with U.S. federal officials while facing legal disputes.