Anthropic Launches Powerful AI Model Mythos, Available for Trial Only to Security Partners

Anthropic has launched Mythos, its most powerful AI model to date, which is currently limited to select security partners and has demonstrated a remarkable ability to uncover thousands of long-standing zero-day vulnerabilities.​

Anthropic Unveils Next-Generation Model "Mythos" for Cybersecurity Tasks

AI powerhouse Anthropic released a preview version of its next-generation "frontier model," Mythos, on Tuesday. Described by the company as its "most powerful" model to date, it is being rolled out initially to a small group of partners through a dedicated safety initiative called Project Glasswing, specifically for cybersecurity-related tasks.

A New Member of the Claude System: Focused on Coding and Reasoning

According to the announcement, Mythos is a general-purpose large-scale model within the Claude ecosystem, boasting exceptional capabilities in programming and logical reasoning. While not trained exclusively for cybersecurity, its primary application during this preview phase is focused on scanning for code vulnerabilities in both first-party and open-source software.

Anthropic reported that in recent internal testing, Mythos identified thousands of zero-day vulnerabilities. Many of these were classified as high-risk, with some issues tracing back to legacy code written one or even two decades ago. This achievement highlights the significant potential of large language models to unearth "historical vulnerabilities" that have remained undetected for years.

Exclusive to Signed Partners, Focusing on "Defensive Security"​

As model capabilities increase, industry concerns regarding potential misuse have also grown. If such tools were exploited by malicious actors, they could drastically accelerate the speed of vulnerability discovery and subsequent attacks. To mitigate these risks, Anthropic explicitly stated that this preview version is not available to the general public and is restricted to signed partners for specific security tasks.

The list of organizations currently participating in the Mythos trial includes dozens of major entities such as Amazon, Apple, Broadcom, Cisco, CrowdStrike, the Linux Foundation, Microsoft, and Palo Alto Networks. These organizations will collaborate on "defensive security efforts" and share their findings upon the project's conclusion to help elevate security standards across the entire technology industry.

Discussions with U.S. Federal Agencies on National-Level Systems

Anthropic also revealed that it is engaged in "ongoing discussions" with U.S. federal officials regarding the potential application of Mythos within critical national systems.

These discussions are taking place alongside existing legal disputes. Previously, the U.S. Department of Defense labeled Anthropic a "supply chain risk entity" after the company refused to allow its technology to be used for autonomous strikes or surveillance targeting U.S. citizens.

Exposure via Accidental Leak

Notably, the existence of Mythos first came to light due to a data security incident. Last month, a draft blog post that had not yet been officially released was mistakenly stored in a publicly accessible data lake. Security researchers subsequently discovered and disseminated the information.

The draft mentioned that Mythos is larger and more intelligent than the previous Opus series. It also contained warnings that, if utilized maliciously, the model could discover and exploit vulnerabilities at a pace that exceeds the ability of defenders to patch them.

Core Takeaways:​

• Mythos is Anthropic's latest frontier AI model, currently restricted to a limited number of security partners for trial.

• The model discovered thousands of zero-day vulnerabilities during testing, showcasing a powerful ability to dig up historical bugs.

• The company is exploring applications for national critical systems with U.S. federal officials while navigating related legal controversies.