A senior U.S. official disclosed Tuesday that Anthropic's advanced artificial intelligence model, Mythos, successfully pinpointed vulnerabilities in America's most sensitive classified computer networks during a controlled exercise, a revelation that intensifies the ongoing debate over AI safety and government oversight.
The official, speaking on condition of anonymity, told The Associated Press that Anthropic collaborated with U.S. intelligence agencies under a program called Project Glasswing. The initiative aims to fortify critical software against potential catastrophic failures that could threaten public safety, national security, and economic stability. Within hours, Mythos identified specific weaknesses, though the official stressed that the model did not necessarily exploit them in that timeframe.
Democratic Senator Mark Warner of Virginia first hinted at the test results during a June 11 hearing before the Senate Committee on Banking, Housing, and Urban Affairs. Warner stated, “This tool broke into almost all of our classified systems, not in weeks but in hours,” attributing the information to General Joshua Rudd, head of the National Security Agency and U.S. Cyber Command. The NSA declined to comment, as did an Anthropic spokesperson.
The revelation comes amid mounting friction between Anthropic and the Trump administration. The California-based AI firm has voiced concerns about military applications of its technology, while the White House has imposed restrictions on some of its models. Earlier this month, the administration ordered Anthropic to block foreign nationals from accessing its latest AI systems, Fable 5 and Mythos 5, citing national security risks. Anthropic complied by disabling the models for all customers, though it disputed the necessity of the directive.
The move followed President Donald Trump's executive order establishing a voluntary framework for federal review of advanced AI systems up to a month before public release. Critics argue the ad hoc approach creates uncertainty, as highlighted in a recent analysis of ad hoc AI regulation under Trump.
A coalition of over 100 cybersecurity executives, including leaders from Adobe and Nvidia, urged the administration to reverse the directive. In a letter, they argued that Anthropic's Mythos models are “quite good” at detecting software flaws but “not uniquely good at these tasks,” and that restricting access to such tools could benefit U.S. adversaries. Many signatories noted they regularly rely on other open-source and foundation models for security audits.
The controversy underscores broader tensions in Washington over AI governance. With congressional oversight increasingly polarized, experts warn that inconsistent policies could leave the nation vulnerable. For context, the divided government threatens to paralyze congressional oversight of emerging technologies.
Anthropic's earlier cooperation with U.S. agencies has also drawn scrutiny. A related report on Anthropic's Fable 5 being punished after cooperation reveals flaws in the current AI review process. Meanwhile, public trust in federal institutions remains at historic lows, as documented by a recent Fox News poll showing record low trust in government.
As the debate intensifies, the White House faces pressure to balance innovation with security. The administration's directive targeting Anthropic's models may set a precedent for how the U.S. regulates AI—a technology that could reshape everything from defense to healthcare.
