Anthropic’s Mythos AI, tested in Project Glasswing with Washington’s intelligence agencies, reportedly identified vulnerabilities in highly sensitive U.S. government computer systems during a sanctioned exercise.

Project Glasswing is a joint initiative by Anthropic to work with tech firms and the government to shield critical software from risks that could threaten public safety, national security, or the economy.

The focus centers on security findings related to a language model and government systems, underscoring potential exposure in handling classified information.

Both the NSA and Anthropic declined to comment publicly on the testing results or any related directive.

Officials indicate the NSA and Anthropic offered no further remarks in their official communications.

There have been conflicting developments this month, including reports that the NSA lost access to Mythos amid the dispute.

The excerpt provides no specific details about the vulnerabilities, the systems involved, or remediation steps.

Anthropic disabled Fable 5 and Mythos 5 for all customers to comply with a government directive, though the company argues the measures were not warranted by the security concerns raised.

Senator Mark Warner cited remarks from the NSA and U.S. Cyber Command leadership suggesting Mythos broke into almost all classified systems within hours, while officials said the model may have detected vulnerabilities without exploiting them.

A U.S. official indicated that vulnerabilities were found within hours but were not exploited during that timeframe.

The administration issued a directive to vet advanced AI systems before public release, following President Trump’s executive order; participation by developers is voluntary.

Industry voices, including over 100 cybersecurity executives from companies like Adobe and Nvidia, urged lifting the directive, arguing Mythos is effective at finding flaws and that restricting access could aid adversaries.