AI model identified government system vulnerabilities in hours during tests

Anthropic’s Mythos AI model identified vulnerabilities in classified U.S. government systems within hours during tests run by intelligence agencies. The tests formed part of Project Glasswing, a collaboration between the company, other firms, and agencies to address security issues.

The Associated Press reported the results this week. The same tests did not establish whether the model could exploit the identified gaps. Officials noted that various parts of the government already use advanced AI tools to detect and fix cyber weaknesses.

A White House official said President Trump has directed steps to safeguard classified systems.

Export controls and company response Earlier this month the administration placed foreign export controls on Mythos and another model called Fable. Anthropic removed both models from public access, stating the action was required to meet the new rules.

The company also sent representatives to Washington to discuss the restrictions with officials. The National Security Agency had been evaluating versions of the tools before the models were taken offline. Anthropic told officials the export limits were unnecessary because the company had already highlighted the security questions.

The firm later proposed closer cooperation with the White House in a submission to the Commerce Department.

Ongoing discussions Talks between the company and administration officials are described as progressing, though no schedule for a permanent resolution has been set. At the G7 Summit last week, President Trump said discussions with the company were going fine. The National Security Agency and Anthropic have not yet responded to requests for additional comment.