Anthropic AI model found flaws in classified U.S. systems, official says

Anthropic’s Mythos artificial-intelligence model found vulnerabilities in highly sensitive U.S. government systems during a controlled testing exercise with intelligence agencies. It identified some flaws within hours, though that did not mean it could exploit them that quickly.

The testing was part of Anthropic’s Project Glasswing, launched on April 7 with Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks. The effort was designed to help defenders find weaknesses in critical software before hostile actors do. Its technical write-up said Mythos Preview could identify and then exploit zero-day vulnerabilities in every major operating system and every major web browser when prompted by a user. Anthropic said the model had already found thousands of high-severity vulnerabilities and committed up to $100 million in usage credits for Mythos Preview across the project, along with $4 million in direct donations to open-source security organizations.

By May 22, Anthropic said roughly 50 partners had used Mythos Preview to find more than 10,000 high- or critical-severity vulnerabilities. On June 2, Anthropic expanded Project Glasswing to about 150 new organizations in more than 15 countries, with many of the partners operating in power, water, healthcare, communications and hardware. Anthropic said most of those partners could face attacks affecting more than 100 million people. On June 9, Anthropic said Claude Mythos 5 was the latest update for Mythos Preview. On June 12, the company said Mythos 5 was temporarily unavailable.

At a June 11 Senate hearing, Sen. Mark Warner said the tool had broken into almost all classified systems in hours, citing the head of the National Security Agency and U.S. Cyber Command, Gen. Joshua M. Rudd. The Senate Intelligence Committee held Rudd’s nomination hearing to become NSA director on January 29.