Key Takeaways
- Anthropic's Mythos AI model identified vulnerabilities in classified U.S. government systems within hours during testing with intelligence agencies
- Senator Warner stated the AI 'broke into almost all of our classified systems, not in weeks but in hours'
- The Trump administration has restricted Anthropic's models over national security concerns, while 100+ cybersecurity experts argue the ban helps adversaries
Anthropic's AI model identified vulnerabilities in highly classified U.S. government computer systems within hours during a joint testing exercise with intelligence agencies, according to an anonymous U.S. official who spoke to the Associated Press. The test, conducted through an Anthropic initiative called Project Glasswing, used the company's Mythos model to probe sensitive government infrastructure.
Senator Mark Warner of Virginia referenced the tests during a June 11 Senate hearing. He said the tool "broke into almost all of our classified systems, not in weeks but in hours," attributing the information to General Joshua Rudd, head of the National Security Agency and U.S. Cyber Command. The NSA declined to comment. Anthropic also declined to comment.
What did Anthropic's Mythos model actually find?
The anonymous official clarified an important distinction: the AI identified vulnerabilities within hours, but that does not mean it exploited them in that timeframe. Finding a flaw is different from weaponizing it. Still, the speed is notable. Traditional penetration testing of classified systems can take weeks of manual effort. An AI model accomplishing the identification phase in hours represents a significant shift in both offensive and defensive capabilities.
Project Glasswing, per the official, brought together tech giants and other companies to secure critical software from what they characterized as "severe" fallout that the Mythos model could pose to public safety, national security, and the economy. The project's existence suggests Anthropic has been working with the government on AI security for some time, though neither party has publicly acknowledged the initiative.
Why is the Trump administration restricting Anthropic's models?
Despite this cooperation, tensions between Anthropic and the Trump administration have escalated. The administration issued a directive earlier this month requiring Anthropic to prevent foreign nationals from using its latest AI models, known as Fable 5 and Mythos 5. Anthropic complied by disabling the models for all customers, though the company stated it did not believe the government's steps were warranted.
The directive followed an executive order Trump signed 10 days earlier, establishing a framework for the federal government to vet national security risks of advanced AI systems for up to a month before public release. Participation by AI developers is voluntary under the order.
Anthropic has also raised concerns about how the U.S. military would use its AI, adding another friction point. The company released Fable widely this month. That model is a limited version of the more advanced Mythos, to which Anthropic has tightly restricted access due to cybersecurity fears.
Cybersecurity experts push back on the restrictions
More than 100 cybersecurity executives and leaders from companies including Adobe and Nvidia sent a letter to the Trump administration asking it to lift the directive. Their argument: the restriction helps U.S. adversaries more than it hurts them.
“Anthropic's Mythos models are 'quite good' at finding flaws in software and weaponizing exploits — but they are 'not uniquely good at these tasks.”
— Letter from 100+ cybersecurity experts to the Trump administration
Many signatories said they regularly use other foundation and open-source models for security audits and training. The letter argued it is dangerous to remove the best cyber defense capabilities "without a good reason" when America's adversaries are rapidly advancing their own AI capabilities. The implication: if U.S. defenders can't use Mythos, they fall behind while foreign actors face no such constraints.
What this means for AI security research
The Mythos model name has not appeared in Anthropic's public documentation. This suggests it's either an internal codename, a classified variant, or a model tier the company has kept private. Anthropic's public-facing models are Claude 3.5 Sonnet, Claude 3 Opus, and others in that family. The existence of a separate model line for security testing raises questions about what other capabilities AI labs are developing outside public view.
The dual-use problem is stark here. The same capability that makes Mythos valuable for finding vulnerabilities makes it dangerous in the wrong hands. Anthropic appears to recognize this, which explains both its government collaboration and its tight access controls. But the administration's restrictions suggest disagreement over who counts as the "right hands."
Logicity's Take
For AI builders, this story highlights a tension you'll face as your models become more capable: the same features that make them valuable for security audits make them dual-use concerns. Anthropic's approach with Project Glasswing, proactively testing against classified systems alongside government agencies, may become a template for responsible deployment. But the government pushback shows that cooperation doesn't guarantee alignment. Teams building agents for code analysis, vulnerability scanning, or infrastructure testing should document their safety processes now. The regulatory environment is shifting fast, and voluntary frameworks could become mandatory.
Frequently Asked Questions
What is Anthropic's Project Glasswing?
Project Glasswing is an Anthropic initiative that brought together tech companies and U.S. intelligence agencies to test AI models against critical software infrastructure. The goal was to identify and secure vulnerabilities before advanced AI models could be misused.
Did Anthropic's AI actually hack government systems?
The AI identified vulnerabilities in classified systems within hours, but according to the anonymous official, this does not mean it exploited those vulnerabilities. Finding a flaw is distinct from successfully weaponizing it.
What is the Mythos AI model?
Mythos appears to be an advanced AI model Anthropic uses for security testing, separate from its public Claude model family. The company has tightly restricted access to Mythos due to cybersecurity concerns.
Why did the Trump administration restrict Anthropic's AI models?
The administration issued a directive requiring Anthropic to prevent foreign nationals from using its latest models, citing national security concerns. This followed an executive order establishing a framework to vet AI systems before public release.
Are other AI models capable of the same security testing?
According to a letter from 100+ cybersecurity experts, Anthropic's models are effective at finding software flaws but are 'not uniquely good at these tasks.' Many experts use other foundation and open-source models for security audits.
More context on how Anthropic's models are being deployed in production workflows
Need Help Implementing This?
Building AI agents for security testing or code analysis? Logicity can help you navigate the technical and compliance landscape. Contact our team for guidance on responsible AI deployment in sensitive contexts.
Source: Fast Company / Associated Press
Huma Shazia
Senior AI & Tech Writer
Produced with AI assistance and reviewed by the Logicity editorial team. Learn more in our Editorial Policy.
