Anthropic launches Project Glasswing with a private AI model for cybersecurity

In short: Anthropic is starting Project Glasswing, a partnership that uses a private AI model called Claude Mythos Preview to help organizations find security weaknesses in their computer systems.

What happened

Anthropic announced Project Glasswing, a cybersecurity program it is launching with partners including Nvidia, Google, Amazon Web Services, Apple, Microsoft, and others. The goal is to help large organizations, and possibly government agencies, spot “vulnerabilities,” which are weak points in software that hackers can take advantage of.

Partners will get access to Claude Mythos Preview, a new AI model that Anthropic says it is not planning to release to the public right now. Anthropic says it is limiting access because the same tool could also help attackers find weak spots.

Anthropic says Mythos Preview can scan systems and flag serious issues with little human help. In its blog post, the company said the model recently found “thousands of high-severity vulnerabilities,” including some affecting major operating systems and web browsers. Anthropic also said the model identified issues and created many related “exploits” (step by step methods to break in) on its own, without human guidance.

The model’s existence became public last month after a data leak that Anthropic attributed to human error. Anthropic said the leak was not caused by a software security flaw.

Anthropic says it will subsidize early use by committing up to $100 million in usage credits. It also plans $4 million in donations to the Linux Foundation and the Apache Software Foundation.

Why it matters

Many everyday services rely on software that is constantly being attacked, like banks, phones, and web browsers. Tools that can quickly find weak points can help defenders patch problems sooner, but they also raise questions about control and oversight, since the same “digital lock pick” could be misused if it spreads.

Source: The Verge AI