Anthropic recently introduced a preview of its latest advanced model, Mythos, targeting a select group of partner organizations for cybersecurity purposes. In a leaked memo, the AI company described Mythos as one of its most powerful models to date.
As part of Project Glasswing, over 40 partners will utilize Mythos for defensive security measures and to enhance software security. Although not specifically developed for cybersecurity, the model will be used to identify vulnerabilities in both proprietary and open-source software.
According to Anthropic, Mythos has detected thousands of critical zero-day vulnerabilities, some dating back one or two decades. Mythos, a general-purpose model, enhances Anthropic’s Claude AI systems with strong coding and reasoning abilities. These advanced models are designed for complex tasks like agent-building and coding.
Organizations participating in previewing Mythos include Amazon, Apple, Broadcom, Cisco, CrowdStrike, the Linux Foundation, Microsoft, and Palo Alto Networks. The aim is to share insights from using the model to benefit the tech industry overall. General availability of the preview is not planned.
Anthropic has been in discussions with federal officials concerning Mythos, despite ongoing legal issues with the Trump administration, as the Pentagon labeled Anthropic a supply-chain risk.
Initial news of Mythos emerged from a security leak reported by Fortune, after a draft blog about the model was found in an unsecured document cache. The leak was attributed to “human error” and detected by security researchers. The leaked document described Capybara—an earlier name for Mythos—as larger and more intelligent than previous models.
Anthropic acknowledged that this model surpasses public models in areas like software coding, reasoning, and cybersecurity, with potential cybersecurity threats if misused by bad actors. The company recently faced issues when 2,000 source code files were mistakenly exposed, leading to a cleanup effort that accidentally took down numerous Github repositories.