On Tuesday, the Associated Press reported that Anthropic, a U.S.-based artificial intelligence (AI) company, uncovered critical security flaws in top-secret government computer systems within hours using its cutting-edge cybersecurity AI model.
According to AP sources, Anthropic partnered with U.S. intelligence agencies to conduct security assessments on highly classified government networks using its Mythos AI model.
This evaluation was conducted as part of Anthropic’s Project Glasswing, a cybersecurity initiative aimed at preemptively addressing potential AI-related security threats by offering the Mythos AI model to organizations for vulnerability detection.
A U.S. official clarified to AP that while Mythos swiftly identified specific vulnerabilities, this doesn’t necessarily mean the model could exploit them within the same timeframe.
During a Senate Intelligence Committee hearing on June 11, Senator Mark Warner (D-Va.) revealed that NSA Director Gen. Joshua Rudd reported Mythos had penetrated nearly all classified systems in hours, not weeks.
The New York Times(NYT) also covered the tests, reporting that NSA analysts formed a red team to evaluate Mythos in a tightly controlled environment.
Red teams typically emulate malicious hackers’ tactics to identify security weaknesses through simulated attacks. The NSA’s red team initiated testing within the agency’s classified systems, which were designed to be accessible only from specific terminals and completely isolated from external networks.
Officials stated that while Mythos rapidly identified security vulnerabilities in the classified network, it did not actually breach the system.
The NYT reported that although the NSA avoided catastrophic outcomes some had feared, analysts were astonished by Mythos’s capabilities, which surpassed expectations in the controlled testing environment.
This announcement coincided with escalating tensions between Anthropic and the Donald Trump administration.
On June 12, citing national security concerns, the U.S. government ordered Anthropic to suspend all foreign access to its Fable5 and Mythos5 models. This restriction applied to foreigners outside the U.S., foreign nationals within the country, and even Anthropic’s non-U.S. employees.
The NYT reported that as the dispute between Anthropic and the administration intensified, the NSA lost access to Anthropic’s AI models.