An advanced artificial intelligence model developed by Anthropic was able to identify vulnerabilities in highly sensitive U.S. government computer systems within hours during a classified testing exercise, according to a U.S. official.
This, once again, highlighted both the extraordinary capabilities and growing security concerns surrounding the latest generation of AI systems.
The disclosure offers one of the clearest indications yet of how rapidly frontier AI models are evolving from productivity tools into technologies capable of performing sophisticated cybersecurity tasks that were once the exclusive domain of highly trained human experts.
Register for Tekedia Mini-MBA edition 20 (June 8 – Sept 5, 2026).
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
According to a U.S. official who spoke to the Associated Press on condition of anonymity, Anthropic worked with American intelligence agencies to evaluate the cybersecurity capabilities of its most advanced AI system, known as Mythos. During the exercise, the model identified vulnerabilities in secure government systems within hours.
The official stressed that identifying vulnerabilities is not the same as exploiting them. While the model was able to locate weaknesses quickly, there is no indication that it independently breached classified systems or gained unauthorized access.
Even so, the findings underscore why governments, intelligence agencies, and technology companies are increasingly treating frontier AI models as dual-use technologies capable of strengthening cyber defenses while also potentially enhancing offensive cyber capabilities.
The testing was conducted through Anthropic’s Project Glasswing initiative, a program designed to assess the risks posed by advanced AI systems and to help secure critical software infrastructure against potentially severe consequences for public safety, national security, and economic stability.
The project brings together government agencies, technology firms, and cybersecurity organizations to evaluate how capable AI models might affect critical infrastructure and digital security. The existence of the testing surfaced publicly earlier this month when Mark Warner referenced the exercise during a Senate Banking Committee hearing.
Warner said, “This tool broke into almost all of our classified systems, not in weeks but in hours.”
The senator attributed the information to Gen. Joshua Rudd, who serves as head of both the National Security Agency and U.S. Cyber Command.
The revelation comes at a time when Anthropic found itself increasingly at odds with the Trump administration over the deployment and export of advanced AI technologies. While the company has cooperated closely with U.S. intelligence and security agencies in assessing cyber risks, it has also expressed concerns about how some of its most powerful models might be used by military and intelligence organizations.
Those tensions escalated earlier this month when President Donald Trump’s administration ordered Anthropic to restrict access to its newest frontier models, Fable 5 and Mythos 5. The directive required the company to prevent foreign nationals from accessing the systems due to concerns about their cybersecurity capabilities.
Anthropic responded by disabling access to the models for all users rather than attempting to implement nationality-based restrictions. The company argued that the government’s response was disproportionate to the risks it had identified and maintained that the security concerns did not justify such sweeping limitations.
The dispute is part of a broader debate unfolding in Washington over how to regulate capable AI systems without undermining America’s technological leadership. Just ten days before the directive was issued, President Trump signed an executive order establishing a framework under which the federal government can evaluate national security risks posed by advanced AI systems before their public release.
Under the framework, participation by AI developers remains voluntary, but it signals a significant expansion of federal oversight of frontier AI models.
The administration’s restrictions have not been universally welcomed within the cybersecurity community. More than 100 cybersecurity experts and executives from major technology companies, including Adobe and Nvidia, have urged the administration to reconsider its approach.
In a letter to government officials, the group acknowledged that Anthropic’s Mythos models are highly effective at identifying software vulnerabilities and generating exploits. However, the experts argued that the systems are “not uniquely good at these tasks” and that comparable capabilities are increasingly available through other commercial and open-source AI models. Many signatories noted that they already rely on multiple foundation models for security audits, vulnerability discovery, and cyber defense training.
The letter warned that limiting access to advanced defensive AI capabilities could inadvertently benefit foreign adversaries.
“It is dangerous to take away the best cyber defense capabilities without a good reason,” the cybersecurity leaders argued, particularly at a time when rival nations are rapidly advancing their own AI programs.
The faceoff between Pentagon and Anthropic was orchestrated by concerns that its newest models could be exploited and used for wrong causes. Advanced AI systems are becoming increasingly valuable tools for detecting vulnerabilities, strengthening cybersecurity, and protecting critical infrastructure. Yet the same capabilities can potentially be used to identify weaknesses that hostile actors might exploit.
Thus, as frontier models continue to improve, policymakers are being forced to confront difficult questions about access, regulation, and national security. The Anthropic case is shaping up to become one of the first major tests of how governments balance those competing priorities.