A powerful artificial intelligence model developed by Anthropic is back in the spotlight after US Senator Mark Warner claimed it breached "almost all" classified government systems during an authorised cybersecurity exercise in just a few hours. The remarks have sparked debate over how advanced AI systems could reshape both cybersecurity and national security.
According to Warner, General Joshua Rudd, who leads the US National Security Agency (NSA) and Cyber Command, briefed lawmakers on an internal red-team exercise involving Anthropic's Mythos AI.
Warner said the model penetrated nearly all targeted classified systems "not in weeks, but in hours." The exercise was reportedly conducted under controlled conditions to test government cyber defences rather than as evidence of a real-world cyberattack.
Mythos is Anthropic's cybersecurity-focused AI model designed to identify software vulnerabilities and assist security teams in detecting weaknesses before attackers can exploit them.
Because of its advanced capabilities, Anthropic has restricted access to the model through its Project Glasswing programme instead of releasing it publicly.
The senator's comments have intensified concerns over how frontier AI models could dramatically accelerate offensive and defensive cybersecurity operations.
However, several cybersecurity experts have cautioned that the reported results came from an authorised internal security test and should not be interpreted as Mythos independently hacking live classified networks.
The claims come days after the US government imposed export restrictions on Anthropic's most advanced AI models, including Mythos, citing national security concerns over their potential misuse. Anthropic has disputed the rationale behind some of the restrictions but has complied with the directive while discussions with US officials continue.
