Anthropic's Project Glasswing: An AI Model to Fight AI Cyberattacks

Anthropic has launched Project Glasswing, a cybersecurity consortium featuring tech giants like Apple, Google, and Microsoft. Participants will use the unreleased Claude Mythos Preview model to find and patch deep-seated vulnerabilities in their software. Anthropic aims to give defenders a head start against future AI-powered cyberattacks by providing early access to this powerful AI.

Anthropic, the parent company of Claude AI, has launched a new cybersecurity initiative called "Project Glasswing." The main goal is to use artificial intelligence to defend against the very cyberattacks that AI is now making possible. To achieve this, the company is also teaming up with rivals in the segment.

Project Glasswing: Anthropic's AI cybersecurity initiative powered by Claude Mythos

The centerpiece of this project is Claude Mythos Preview. Anthropic describes this new AI model as a major leap forward. While it wasn't specifically designed for hacking, its advanced coding and reasoning skills make it a formidable tool for finding security holes. According to Anthropic, the model has already autonomously identified thousands of critical vulnerabilities, including bugs in every major operating system and web browser.

Because of this power, Anthropic is not releasing the model to the public. As reported by The Verge, the company fears that if Mythos Preview fell into the wrong hands, it could become a "meaningful accelerant" for hackers. Instead, they are keeping it behind a "glass wing," giving access only to a select group of defensive partners.

Unlikely allies

Project Glasswing has accomplished something rare in Silicon Valley: it has brought fierce rivals together. The consortium includes over 45 organizations, such as Apple, Google, Microsoft, Amazon Web Services, and NVIDIA. Financial giants like JPMorganChase and infrastructure groups like the Linux Foundation are also on board.

The logic behind this collaboration is defensive. It wants to give these companies private access to the model to find and patch vulnerabilities in their own software before similar capabilities become widely available to malicious actors. Logan Graham, Anthropic's frontier red team lead, told WIRED that we must prepare for a world where these high-level hacking capabilities are common within the next year or two.

A complicated path forward

Despite the high-tech mission, Anthropic's journey hasn't been without friction. The company recently faced tension with the U.S. government after refusing to remove certain guardrails for military use. This even led the company to get a "supply chain risk" designation from the Pentagon. Furthermore, reports suggest that a hacker used an earlier version of Claude against government agencies in Mexico earlier this year.

To support the effort, Anthropic is committing $100 million in usage credits and making significant donations to open-source foundations.