NSA Already Uses Anthropic’s Most Powerful AI Model “Myhtos” Despite Pentagon Ban

Despite an ongoing legal dispute between the US Department of Defense and AI company Anthropic, the American intelligence agency NSA is apparently already using the company’s latest and most powerful model: Claude Mythos Preview. This is reported by two independent sources to the US news outlet Axios. The case sheds light on the growing influence of AI in national security and the contradictions within the US government.

What Is the NSA and Why Is It Relevant?

The National Security Agency is the largest and most technically advanced intelligence service in the United States. It falls under the Department of Defense (recently renamed the Department of War) and is primarily responsible for the collection, analysis, and protection of signals and digital communications — known as Signals Intelligence (SIGINT). With tens of thousands of employees and one of the largest supercomputer networks in the world, the NSA is considered one of the most powerful cyber intelligence organizations globally.

That is precisely what makes this report explosive: the Department of Defense, to which the NSA is subordinate, officially designates Anthropic as a “Supply Chain Risk” — that is, a risk to the supply chain of security-critical technologies. At the same time, one of its subordinate agencies is using exactly the tools of that very company. In February, the Department had attempted to exclude Anthropic as a supplier and to encourage its contractors to follow suit. That legal dispute has not yet been resolved.

The Contradiction: Security Risk and Indispensable Tool at Once

The situation within the US government is divided. On one side stand defense officials who accuse Anthropic of being unreliable because the company refuses certain use cases. Specifically, the Pentagon had demanded the right to use the Claude model for “all lawful purposes.” Anthropic, however, insisted on excluding mass domestic surveillance and the development of autonomous weapons.

On the other side, parts of the government are pushing to end the conflict in order to retain access to the technologically leading tools. Anthropic CEO Dario Amodei therefore met with Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent at the White House on April 17, 2026, to discuss the use of Mythos in government agencies as well as the company’s security practices. Both sides described the meeting as productive.

“The work to defend the world’s cyber infrastructure could take years; the capabilities of frontier AI will likely advance considerably in the coming months. For cyber defenders to maintain the upper hand, we must act now.” (Anthropic)

What Can Claude Mythos Preview Do?

According to Anthropic, Claude Mythos Preview is the company’s most capable model to date and, by its own account, surpasses the abilities of all but the best human security experts in the field of cybersecurity. The model is capable of identifying security vulnerabilities in software nearly fully autonomously and developing working exploits without human guidance.

During internal testing, Mythos discovered thousands of critical vulnerabilities across all major operating systems and web browsers. Three specific findings were made public:

• A 27-year-old vulnerability in OpenBSD, an operating system considered particularly secure for firewalls and critical infrastructure. The flaw made it possible to crash affected machines using only a network connection.
• A 16-year-old vulnerability in FFmpeg, a widely used video processing library. Automated testing tools had already executed the relevant line of code five million times without detecting the problem.
• Multiple linked vulnerabilities in the Linux kernel, whose combination allowed an attacker to escalate from ordinary user access to full control over a machine.

All of the mentioned vulnerabilities were reported to the responsible software maintainers and have since been patched. Additional discovered vulnerabilities have initially been published by Anthropic only as cryptographic hashes, with full details to be disclosed only after a fix is in place.

Why Mythos Is Not Publicly Available

Due to its extraordinary offensive capabilities, Anthropic has decided not to make Claude Mythos Preview generally accessible. Access is restricted to around 40 organizations that have committed to responsible use. The company’s concern: in the wrong hands, the model’s capabilities could have devastating consequences for the economy, public safety, and national security.

In the long term, Anthropic plans to make models of this performance class available safely and at scale. To do so, suitable protective mechanisms must first be developed that can detect and block dangerous outputs. These safety systems are to be tested first on a less risky model — an upcoming Claude Opus model.

Project Glasswing: Who Already Has Access

To deploy Mythos’s capabilities specifically for defensive purposes, Anthropic has launched the initiative Project Glasswing. The goal is to use the model for defensive security work and to share insights gained with the entire industry. Anthropic is providing up to $100 million in usage credits as well as $4 million in direct donations to open-source security organizations.

The officially confirmed founding partners of Project Glasswing include twelve prominent companies and organizations:

• Amazon Web Services
• Anthropic
• Apple
• Broadcom
• Cisco
• CrowdStrike
• Google
• JPMorganChase
• Linux Foundation
• Microsoft
• NVIDIA
• Palo Alto Networks

In addition, more than 40 further organizations that develop or operate critical software infrastructure are being granted access. According to one source, the NSA is among the agencies with access to the model that have not been publicly named. The British intelligence service MI5 is also said to have gained access through the United Kingdom’s AI Security Institute.

Outlook: AI as the Key to Cyber Defense

The case illustrates how much powerful AI models have already moved to the center of state security strategies today. The NSA is apparently using Mythos primarily to scan its own infrastructure for vulnerabilities — similar to other organizations with access to the model. Whether and how the conflict between the Pentagon and Anthropic will be resolved remains open. What is clear, however, is this: the technological appeal of Claude Mythos appears to be stronger than political differences.