Anthropic unveils Mythos: Powerful new model focused on cybersecurity
Anthropic has officially launched a preview of its newest and most powerful frontier model, Mythos. The AI will initially be utilized by a select group of major companies with the primary goal of engaging in defensive cybersecurity work and securing critical software.
Context
Mythos, which previously leaked under the codename “Capybara” in a data security incident last month, is described as the most sophisticated and highest-performing model ever developed by Anthropic. Accidental leaks already anticipated that the new tier of model would be larger and more intelligent than its Opus models. Now, the information has been officially introduced through an initiative dubbed “Project Glasswing,” pointing the model’s abilities towards scanning and fixing vulnerabilities in both first-party and open source software systems.
Details
For now, Anthropic will not make the model generally available to the public to prevent potential weaponization by bad actors. Behind the “Project Glasswing” initiative are 12 partner organizations, including Amazon, Apple, Broadcom, Cisco, CrowdStrike, the Linux Foundation, Microsoft, and Palo Alto Networks. Specifically, another 40 organizations will also gain special access to the preview.
Anthropic claims that the model’s advanced “reasoning” capabilities and strong agentic coding skills have quickly proven their immense value: over the past few weeks, Mythos proactively identified thousands of zero-day vulnerabilities (many of them critical and one to two decades old). The goal is for these partner organizations to use the tool to bolster their defenses, and eventually share what they’ve learned to benefit the tech industry as a whole.