Anthropic’s Mythos AI Found Flaws in Classified US Systems Within Hours, Officials Say
Testing Exercise Reveals Rapid Vulnerability Detection
Poinews.com – A US official disclosed that Anthropic’s AI model discovered weaknesses in highly secure government systems during a recent evaluation. The model, part of the company’s Mythos series, was tested in partnership with intelligence agencies, according to the source. While the AI pinpointed specific vulnerabilities quickly, it did not necessarily have the capability to exploit them in the same timeframe.
Collaboration and Project Glasswing
The security assessments were conducted under Anthropic’s Project Glasswing, an initiative designed to evaluate advanced AI systems for potential risks. This project aimed to safeguard critical software from significant threats, particularly those that could impact public safety, national security, and economic stability. Tech firms joined the effort to identify and address vulnerabilities before they could be exploited.
Senator’s Statement on AI Capabilities
During a June 11 Senate hearing, Democratic Senator Mark Warner from Virginia highlighted the findings. He stated,
“This tool breached almost all of our classified systems, not in weeks but in hours,”
referencing information provided by General Joshua Rudd, head of the National Security Agency (NSA) and US Cyber Command.
Tensions with Trump Administration
Anthropic has faced increasing friction with the Trump administration over AI usage. The company expressed concerns about the military’s potential applications of its models, while the administration sought to limit their use. Recently, a directive was issued requiring Anthropic to block foreign nationals from accessing its Fable 5 and Mythos 5 models.
New Directive on Model Access
The order followed an executive directive signed by President Donald Trump, which mandated a one-month review of national security risks for advanced AI systems before public release. Developers were given the option to participate voluntarily. Anthropic complied by disabling the models for all customers, though it questioned the necessity of the restrictions.
Cybersecurity Experts’ Concerns
Over 100 cybersecurity professionals, including leaders from Adobe and Nvidia, have urged the government to reconsider the directive. They argue that while Mythos models are effective at detecting software flaws, they are not the only tools capable of such tasks. Many rely on other models for audits and training, and they warn that removing top cyber defense capabilities without justification could weaken America’s security posture.
