Anthropic’s AI model, Mythos, has uncovered significant vulnerabilities in sensitive U.S. government computer systems during a testing exercise, according to reports from Associated Press and Reuters.
In collaboration with Washington intelligence agencies, Anthropic conducted tests using Mythos as part of the Glasswing Project, a restricted program aimed at identifying and resolving critical software vulnerabilities before they can be exploited by attackers.
During a recent Congressional hearing, Senator Mark Warner from Virginia disclosed that he was informed by NSA chief Joshua Rudd that Mythos had infiltrated nearly all classified systems within hours, rather than weeks.
While an unidentified U.S. official noted that Mythos identified certain vulnerabilities quickly, this does not imply that the model was capable of exploiting them within that timeframe.
Neither the White House, Anthropic, nor the Department of Defense has responded to Reuters’ requests for comment on this issue.
The relationship between Anthropic, which is preparing for an initial public offering, and the U.S. government has been fraught with tension. The company has previously declined to allow the U.S. military to use its AI models for domestic surveillance and fully autonomous weapon systems, leading to the government’s response of placing it on a national security blacklist.
Additionally, this month, the U.S. government ordered Anthropic to suspend the export of its latest AI models, Mythos and Fable, to destinations worldwide and to all foreign nationals, citing national security concerns.
Earlier reports from The New York Times indicated that the NSA had lost access to Mythos amid these disputes.
