Anthropic's artificial intelligence hacked the U.S. NSA's systems in a matter of hours

Anthropic’s Mythos 5 artificial intelligence managed to hack into the U.S. National Security Agency’s systems in just a few hours during testing. The latest model demonstrated unprecedented capabilities for hacking attacks, gaining access to nearly all of the agency’s classified systems. 

The Economist reported this, citing a statement by Joshua Rudd, director of the NSA and head of the Pentagon’s Cyber Command. According to Mark Warner, vice chairman of the Senate Intelligence Committee, the system demonstrated capabilities that would have taken conventional cybersecurity experts years to achieve, although the exact level of access it gained has not yet been disclosed.

Anthropic unveiled the Mythos 5 model on April 7, 2026, claiming it could detect and exploit critical vulnerabilities in operating systems, web browsers, and software. According to the U.S. publication Forbes, many of the identified flaws had existed for decades and had gone unnoticed during numerous previous audits. The model was tested as part of “Project Glasswing,” where it helped identify more than 10,000 high- and critical-level vulnerabilities. Due to the risk of the system becoming a dangerous tool for hackers and foreign intelligence agencies, Anthropic’s management decided against a public launch and granted access to the technology only to a limited circle of partners.