Anthropic’s AI model Mythos discovered security vulnerabilities in highly sensitive US government computer systems during tests under the classified Project Glasswing initiative [1, 2, 3, 4, 5]. Mythos found these flaws in hours, a notably faster pace compared to the typical weeks it takes to identify such issues [1, 2, 3, 4, 5].
Senator Mark Warner cited NSA chief Joshua Rudd during a June 11, 2026, congressional hearing, saying, “Mythos broke into almost all of our classified systems, not in weeks, but in hours” [3, 2, 4]. However, an unnamed US official clarified that although Mythos quickly identified vulnerabilities, it did not mean the model could exploit them within the same timeframe [1].
Project Glasswing is a secretive government effort aimed at discovering and fixing critical software vulnerabilities before adversaries can exploit them [1, 3, 5]. In April 2026, Mythos access was initially granted to 50 organizations for cybersecurity testing related to this project [6].
The relationships between Anthropic and the US government have been tense recently. The government ordered Anthropic in June 2026 to suspend exports of its Mythos and Fable AI models internationally and to foreign nationals due to national security concerns [1, 3, 4, 5]. Around the same time, the NSA reportedly lost access to Mythos amid ongoing disputes with Anthropic [1, 4].
Anthropic has refused US military requests to use its AI models for domestic surveillance and fully autonomous weapons systems [1, 4, 5]. CEO Dario Amodei said the company is “working with the US government to balance national security and AI innovation” and advocates “for international AI safety cooperation” [5].
The rapid ability of advanced AI models like Mythos to detect cybersecurity vulnerabilities has sparked renewed debates in Washington about the risks and regulation of AI technology [5, 6]. Experts note that similar capabilities exist in open-source AI models, which may increase risks if these tools fall into malicious hands [6].
The next scheduled steps include ongoing government oversight and possible further restrictions on AI model exports, as national security concerns remain at the forefront.
