A top-tier artificial intelligence model developed by Anthropic has uncovered hidden security vulnerabilities in highly secure classified computer systems of the U.S. government, according to an anonymous U.S. official speaking to the Associated Press.

The AI firm has partnered with U.S. intelligence agencies to run confidential cybersecurity tests with its self-developed Mythos model. The AI system is capable of spotting undiscovered system weaknesses in an extremely short period of time. Still, the official made a clear clarification: fast identification of loopholes does not mean the AI can take advantage of these security flaws to launch real cyber attacks in the same span.

All the cybersecurity evaluations are part of Anthropic’s Project Glasswing. This cross-industry program gathers leading technology companies to upgrade protection for critical global software systems. The core purpose of the project is to respond to potential huge threats brought by advanced AI tools. If such powerful AI technology is abused illegally, it will bring devastating impacts on national security, public safety and economic stability.

The surprising testing result has been mentioned publicly by senior U.S. lawmaker Mark Warner. The Democratic senator pointed out that this advanced AI tool can break through nearly all American classified computer systems rapidly, while traditional manual detection methods require far longer working hours. He noted that this authoritative conclusion comes from General Joshua Rudd, the chief of both the National Security Agency and U.S. Cyber Command.

This incident sparks widespread discussions about the dual nature of modern advanced AI. On one hand, AI can help government departments detect cyber risks in advance and optimize digital defense systems effectively. On the other hand, its extraordinary capability to scan system loopholes also arouses growing concerns over AI risks. Governments around the world are now attaching greater importance to standardized supervision of high-performance AI models, to stop the technology from being used for illegal cyber activities.