Claude Mythos Preview: Anthropic Revolutionizes Security

A Major Breakthrough with Claude Mythos Preview

Anthropic recently unveiled Project Glasswing, an initiative bringing together eleven major names in the tech industry around Claude Mythos Preview. This artificial intelligence model stands out for its ability to identify security vulnerabilities that neither human audits nor automated tests had previously detected.

The announcement of this technology has generated significant attention. Anthropic has indeed developed an AI model capable of spotting thousands of critical vulnerabilities in the most widely used software worldwide. This includes all major operating systems as well as web browsers. The company has chosen to deploy this technology among large enterprises to prevent it from being exploited for malicious purposes.

Vulnerabilities Ignored for Decades

The Claude Mythos Preview model has been specifically designed to analyze code and identify vulnerabilities. In tests on the CyberGym benchmark, which precisely evaluates these skills, Mythos Preview achieved an impressive score of 83.1%, far surpassing Anthropic's previous model, Claude Opus 4.6, which scored 66.6%.

The examples provided by Anthropic clearly illustrate the effectiveness of this model. For instance, a flaw present for 27 years in the OpenBSD system was discovered. This vulnerability allowed an attacker to remotely crash any connected machine. Additionally, Mythos Preview identified an issue in FFmpeg, a software library used in thousands of applications for reading and encoding videos. This flaw had evaded five million automated tests before being detected.

Another case, perhaps the most concerning, involves the Linux kernel. Mythos Preview successfully identified and combined several vulnerabilities that allowed for escalation from ordinary user access to full control of the machine. Detecting this chain of flaws typically requires genuine human expertise. Anthropic notes that all these vulnerabilities have been reported to the relevant teams and have since been fixed.

A Massive Collaboration for Enhanced Security

Anthropic is not working alone on this ambitious project. Project Glasswing brings together major players such as:

• AWS
• Apple
• Broadcom
• Cisco
• CrowdStrike
• Google
• JPMorganChase
• the Linux Foundation
• Microsoft
• NVIDIA
• Palo Alto Networks

This project provides shared access to the model for their defensive security efforts. Furthermore, over 40 additional organizations, including many open-source projects, also benefit from this access.

The funding matches the ambitious goals set forth. Anthropic is investing up to $100 million in usage credits for participants. Additionally, four million dollars are directly allocated to open-source security organizations, with $2.5 million via the Linux Foundation for Alpha-Omega and OpenSSF, and $1.5 million to the Apache Software Foundation. This is due to the fact that a large portion of the software that powers the internet relies on open-source projects, often without dedicated security resources.

It was recently reported that Claude Opus 4.6 detected over a hundred bugs within Firefox in just two weeks. The results achieved with this new AI are even more impressive.

Restricted Access for Security Reasons

For security reasons, Claude Mythos Preview will not be made available to the general public. After the "research preview" phase, access will be charged at $25 for one million tokens in input and $125 for output. Anthropic is also in discussions with representatives from the U.S. government regarding the offensive and defensive implications of this model. The company is considering the creation of an independent body to coordinate these efforts across the industry.