Anthropic's 'Mythos' AI Discovers Vulnerabilities in U.S. Government Systems in Hours

Anthropic's artificial intelligence (AI) model 'Mythos,' known for its expert-level cybersecurity capabilities, has reportedly identified vulnerabilities in classified U.S. government systems in just a few hours.

The Associated Press reported on June 23 (local time), citing an anonymous official, that in recent tests conducted in collaboration with U.S. intelligence agencies, Mythos identified vulnerabilities in highly sensitive and secure government computer systems.

However, the official added that Mythos was not yet at a level where it could exploit those security vulnerabilities within that timeframe.

This test was conducted as part of 'Project Glasswing,' a security initiative Anthropic is carrying out with trusted institutions and companies.

Previously, Senator Mark Warner (Democrat-Virginia) also mentioned the test during a congressional hearing on June 11, stating, "This tool penetrated almost all of our classified systems in just a few hours, not weeks."

Senator Warner stated that he received this information from General Joshua Rudd, commander of the National Security Agency (NSA) and U.S. Cyber Command, but the NSA declined to comment on the matter.

Anthropic has been at odds with the federal government, including being designated as a 'supply chain risk' company following conflicts with the U.S. Department of Defense over the scope of its AI usage.

In particular, shortly after Anthropic released its Mythos 5 model and the Fable 5 model—which includes additional safety guardrails—to the public, the U.S. government issued export control guidelines requiring that foreign nationals be blocked from accessing these models, citing national security concerns.

Consequently, Anthropic, which lacked the functionality to verify users' nationalities, blocked all user access to both models.

More than 100 cybersecurity experts and leaders from major tech companies, including Adobe and Nvidia, have sent a letter requesting that the U.S. government withdraw these control guidelines on Anthropic's models, arguing that they could actually benefit adversarial nations.

(Photo: Yonhap News)
※ Please note: This article was translated by AI and may contain errors.