Introducing Mythos AI
Anthropic has introduced a preview of its cutting-edge AI model, named Mythos, designed to tackle complex challenges. This new frontier model is being
made available to a select group of partner organizations, specifically for their work in cybersecurity. Described by the AI startup as one of their "most powerful" creations to date, Mythos represents a significant leap forward in AI development. Its limited debut is a strategic component of a new security initiative called Project Glasswing. Through this project, twelve chosen partner organizations will deploy Mythos to enhance defensive security measures and safeguard critical software infrastructure. Although not exclusively trained for cybersecurity tasks, the model's robust capabilities enable it to meticulously scan both proprietary and open-source software systems, effectively identifying potential code vulnerabilities. Anthropic has reported that over recent weeks, Mythos has successfully pinpointed thousands of zero-day vulnerabilities, many of which are classified as critical. Intriguingly, a substantial number of these discovered flaws are remarkably old, some dating back one to two decades, highlighting the model's profound depth in code analysis.
Project Glasswing Initiative
Project Glasswing signifies Anthropic's strategic approach to leveraging its advanced AI for enhanced security. This initiative involves an exclusive cohort of twelve prominent organizations, including giants like Amazon, Apple, Broadcom, Cisco, CrowdStrike, the Linux Foundation, Microsoft, and Palo Alto Networks. These partners are granted early access to the Mythos model, empowering them to utilize its formidable capabilities for "defensive security work." The primary objective is to bolster the security posture of critical software and identify latent vulnerabilities. A key aspect of Project Glasswing is the collaborative sharing of insights; the participating organizations will contribute their findings and experiences with Mythos back to the broader tech industry. This collective knowledge-sharing aims to accelerate the development of more secure software ecosystems for everyone. While this preview is not accessible to the general public, an additional forty organizations will receive access to the Mythos preview outside of the formal partnership, broadening its impact and potential for security improvements across various sectors.
Mythos Capabilities & Leak
Mythos is engineered as a versatile, general-purpose model within Anthropic's Claude AI systems, boasting exceptional agentic coding and reasoning proficiencies. Frontier models from Anthropic are recognized as their most advanced, intended for highly complex tasks such as building AI agents and sophisticated coding projects. The initial news of Mythos's existence surfaced unexpectedly through a data security incident. A draft document detailing the model, then referred to as "Capybara," was inadvertently left exposed in an unsecured data cache, accessible via a public data lake. Anthropic attributed this oversight to "human error," with security researchers being the first to discover the leak. The leaked document described "Capybara" as a new tier of model, substantially larger and more intelligent than their existing Opus models, and unequivocally declared it "by far the most powerful AI model we’ve ever developed." In that leak, Anthropic indicated that its new model significantly outperformed current public models in areas such as software coding, academic reasoning, and cybersecurity. This foresight also acknowledged the potential for such a powerful model to be misused by malicious actors to discover and exploit software bugs, rather than using it for their intended constructive purposes.
