Anthropic releases preview of new AI model Mythos for cyber defence

Anthropic, an American artificial intelligence (AI) company, released a preview of its next-generation frontier AI model called Mythos. The new model will be deployed as part of the company’s new cybersecurity initiative, “Project Glasswing,” which brings together the world's most powerful technology companies to use the AI model for defensive security work. 

Anthropic, in a company blog post, revealed that Mythos will be used to identify vulnerabilities across critical software systems. “Mythos Preview has already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser,” the company added.

What is Project Glasswing?

Project Glasswing is a new cybersecurity initiative that brings together Amazon Web Services (AWS), Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks in an effort to secure critical software. 

These companies and over 40 additional organisations will be using Mythos Preview to secure both first-party and open-source systems, and will share what they have learned for the world’s benefit. 

“No one organisation can solve these cybersecurity problems alone,” Anthropic said.

“Frontier AI developers, other software companies, security researchers, open-source maintainers, and governments across the world all have essential roles to play,” it added.

Driving the initiative is Claude Mythos Preview, Anthropic’s cybersecurity-focused AI model. The term “Mythos,” rooted in Ancient Greek, signals the ability to decode complex data and code patterns.

Claude Mythos Preview: Features and performance benchmarks

As per the blog post, Mythos Preview is designed to detect vulnerabilities across software systems and previously unknown zero-day vulnerabilities. It supports critical cybersecurity workflows such as penetration testing, binary analysis, and black-box testing of systems. 

In addition, it seamlessly integrates enterprise and open-source security processes to help secure endpoints and infrastructure components. It also offers automation in vulnerability detection and triage, strengthening secure software development.

The preview model has already spotted vulnerabilities across platforms, including operating systems and web browsers. The company also shared a couple of examples of its work, such as spotting a 27-year-old vulnerability in OpenBSD, a 16-year-old vulnerability in FFmpeg, and several chained vulnerabilities in the Linux kernel.

In terms of benchmark performance, Mythos Preview scored 83.1% on the CyberGym benchmark for cybersecurity vulnerability reproduction, compared to 66.6% for Opus 4.6. On coding benchmarks, the gap is even wider; on SWE-bench Verified, Mythos Preview scored 93.9% against Opus 4.6's 80.8%. In the SWE-bench Multilingual, Mythos Preview scored 59% in comparison to Opus 4.6’s 27.1%.

Claude Mythos Preview: Availability 

Anthropic will be offering up to $100 million in model usage credits to support Project Glasswing participants. In addition, it will be available on platforms such as the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. However, the preview has not been released for public use. Therefore, access is restricted to selected partners of Project Glasswing.

However, a security professional can request access via the Cyber Verification Program and Claude for the Open Source initiative.