Mythos: Anthropic’s Next-Generation AI for Cybersecurity

Mythos is an innovative AI model introduced by Anthropic, an American artificial intelligence company. This model is set to play a crucial role in the company’s latest cybersecurity initiative known as Project Glasswing. This project aims to unite the world’s leading technology firms to employ the AI model for defensive security purposes.

Understanding Project Glasswing

Project Glasswing represents a significant effort in cybersecurity, collaborating with major players such as Amazon Web Services (AWS), Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks to enhance the security of essential software.

Microsoft is joined by more than 40 other organisations that will leverage Mythos to protect both proprietary and open-source systems. Insights gained from this collaboration will be shared for the wider benefit of the global community.

Mythos: A Powerful Tool for Cybersecurity

Project Glasswing is described as an urgent initiative focused on securing the world’s most critical software. The AI model, Claude Mythos Preview, is designed to identify software vulnerabilities more effectively than most human experts.

According to Anthropic, no single organisation can tackle the challenges in cybersecurity independently. In combination, frontier AI developers, various software companies, security researchers, open-source maintainers, and governments all play crucial roles in this integral initiative.

Claude Mythos Preview: Key Features and Performance Metrics

Mythos Preview is specifically crafted to uncover vulnerabilities throughout software systems, including unknown zero-day vulnerabilities, and supports essential cybersecurity tasks like penetration testing, binary analysis, and black-box testing.

Moreover, it facilitates the integration of both enterprise and open-source security processes, enhancing the protection of endpoints and infrastructure. Its automation capabilities streamline the tasks of vulnerability detection and prioritisation, bolstering the development of secure software.

The preview model has discovered vulnerabilities across various platforms, highlighting its efficacy. Notable instances include the identification of a 27-year-old vulnerability in OpenBSD, a 16-year-old vulnerability in FFmpeg, and multiple chained vulnerabilities within the Linux kernel.

On benchmark performance, Mythos Preview achieved a score of 83.1% on the CyberGym benchmark for reproducing cybersecurity vulnerabilities, surpassing Opus 4.6’s score of 66.6%. In coding benchmarks, the disparity is significant: Mythos Preview scored 93.9% on the SWE-bench Verified, compared to Opus 4.6’s 80.8%. Additionally, in the SWE-bench Multilingual, Mythos Preview scored 59%, in contrast to the 27.1% of Opus 4.6.

Claude Mythos Preview: Availability

To support participants in Project Glasswing, Anthropic will allocate up to $100 million in usage credits for the model. Additionally, it will be accessible on platforms including the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. However, the preview remains unavailable to the general public, limiting access to selected partners of Project Glasswing.

Security professionals can request access through the Cyber Verification Program and the Claude for Open Source initiative.