New Delhi, June 24: Anthropic’s advanced AI model Mythos reportedly identified vulnerabilities in highly sensitive U.S. government computer systems during a restricted testing exercise conducted with American intelligence agencies, according to a media report.
The testing was carried out under a programme known as Project Glasswing, which is designed to detect and help address weaknesses in critical software systems before they can be exploited by malicious actors. The exercise reportedly involved using Mythos to scan classified government infrastructure for security flaws.
According to the report, the model was able to identify vulnerabilities in several systems within a matter of hours. However, officials cited in the report indicated that detecting weaknesses did not necessarily mean the AI system had successfully exploited those systems in the same timeframe.
The development has drawn attention because it highlights the growing role of advanced AI models in cybersecurity and national security applications. Models like Mythos are designed to analyze code, identify flaws, and accelerate vulnerability detection, potentially giving governments and organizations a faster way to secure critical infrastructure.
The reported testing also comes amid a complicated relationship between Anthropic and the U.S. government. In recent months, the company has faced restrictions related to the use and export of some of its most advanced AI systems, particularly over national security concerns and disagreements over military applications.
Reports indicate that Anthropic had previously resisted allowing certain uses of its models in areas such as domestic surveillance and fully autonomous weapons systems. At the same time, the U.S. government has shown strong interest in the company’s technology for cybersecurity and defense-related purposes.
The latest report underscores the increasingly strategic importance of frontier AI systems in cyber defense, especially as governments seek tools capable of protecting classified networks and critical software from sophisticated attacks.
Neither Anthropic nor U.S. government agencies had publicly detailed the full scope of the reported testing exercise at the time of reporting. Still, the episode is likely to intensify the broader debate around how advanced AI models should be governed, deployed, and secured when used in national security environments.
Author
About the Author
