The Associated Press reported that Anthropic’s Mythos model identified vulnerabilities in highly sensitive U.S. government systems during Project Glasswing tests. Senator Mark Warner said NSA director Joshua Rudd told him Mythos "broke into almost all of our classified systems" within hours, though AP cautioned detection did not necessarily mean exploitation. The episode comes amid export restrictions and a broader dispute between Anthropic and U.S. authorities.
AP: Anthropic’s Mythos AI Found Vulnerabilities In U.S. Classified Systems During Project Glasswing Tests
FILE PHOTO: Illustration shows Anthropic logo
June 23 — The Associated Press reported that Anthropic’s large language model Mythos identified weaknesses in highly sensitive U.S. government computer systems during a controlled testing exercise.
Anthropic worked with U.S. intelligence agencies under a restricted initiative called Project Glasswing, a program designed to discover and remediate critical software flaws before malicious actors can exploit them.
At a congressional hearing this month, Senator Mark Warner (D‑Va.) said he was told by National Security Agency director Joshua Rudd that Mythos "broke into almost all of our classified systems, not in weeks, but in hours." The AP, citing an unnamed U.S. official, added a caution: while the model flagged certain vulnerabilities within hours, that does not necessarily mean it could exploit them in that same timeframe.
Sen. Mark Warner: "Mythos broke into almost all of our classified systems, not in weeks, but in hours."
Requests for comment from Reuters to the White House, Anthropic and the Department of Defense received no immediate replies, the AP said.
Context And Implications
Anthropic, which is preparing for an initial public offering, has had a tense relationship with the U.S. government. The company declined to allow its AI models to be used by the U.S. military for domestic surveillance and fully autonomous weapons, and was placed on a national security blacklist. This month, the U.S. ordered a suspension of exports of Anthropic’s latest Mythos and Fable models to foreign destinations and to foreign nationals, citing national security concerns. The New York Times also reported the NSA had lost access to Mythos amid the dispute.
Security researchers and policymakers will likely scrutinize the distinction between an AI model’s ability to identify vulnerabilities and its ability to exploit them. The episode underscores growing interest in using advanced AI for defensive cyber exercises while highlighting policy tensions over access and control.
(Reporting by Anusha Shah in Bengaluru; Editing by Muralikumar Anantharaman and Edwina Gibbs)
Help us improve.
