One of Anthropic's most advanced artificial intelligence models identified vulnerabilities inside highly sensitive U.S. government computer systems during a controlled testing exercise conducted with intelligence agencies, according to a new report.
The official, who spoke on condition of anonymity, said Anthropic's Mythos model discovered weaknesses in classified systems within hours during tests carried out under Project Glasswing, an initiative aimed at strengthening critical software infrastructure, CNBC noted.
The official said the model identified vulnerabilities but did not necessarily exploit them. Project Glasswing brought together Anthropic, intelligence agencies and technology companies to evaluate risks that advanced AI systems could pose to public safety, national security and the broader economy.
Sen. Mark Warner of Virginia publicly referenced the exercise during a June 11 hearing before the Senate Committee on Banking, Housing and Urban Affairs. Warner said he had been informed by Gen. Joshua Rudd, who leads both the National Security Agency and U.S. Cyber Command, that the system had "broken into almost all of our classified systems, not in weeks but in hours." Reuters later reported that government officials clarified the exercise involved identifying vulnerabilities rather than carrying out attacks on the systems.
Concerns over cybersecurity and digital warfare have intensified following years of Russian cyber activity tied to the war in Ukraine and rising tensions between the United States and China over technology and national security. U.S. officials have repeatedly warned about threats to critical infrastructure, including energy grids, telecommunications networks and government systems.
Anthropic's relationship with the Trump administration has become increasingly strained. The administration recently restricted access to the company's Fable 5 and Mythos 5 models over national security concerns. Anthropic said it disabled access to the models for all customers in order to comply with a government directive aimed at preventing foreign nationals from using them.
The restrictions followed an executive order signed by President Donald Trump earlier this month establishing a framework for federal agencies to evaluate the national security implications of advanced AI systems before they are publicly released. The Washington Post, citing the Associated Press, reported that the order created a voluntary review process lasting up to 30 days.
Anthropic has argued that the government's actions were not justified by the concerns it raised about potential security issues. A group of more than 100 cybersecurity experts and executives from companies including Adobe and Nvidia also urged the administration to reconsider the directive. The group said advanced models such as Mythos are highly effective at detecting software flaws but are not uniquely capable compared with other tools, according to Reuters.
Many signatories to the letter said they routinely use a variety of foundation models and open-source systems for security testing and training. They warned that limiting access to advanced defensive tools without sufficient justification could undermine efforts to protect networks while rivals continue to advance their own cyber capabilities.
