The Pentagon is simultaneously relying on Anthropic's most powerful AI model to defend American government networks and racing to cut all ties with the company that built it. The contradiction is now official policy.

On Tuesday, the Defense Department's chief technology officer Emil Michael confirmed that the Pentagon is deploying Anthropic's Claude Mythos Preview model to find and patch software vulnerabilities across US government systems. The deployment is happening through Project Glasswing, Anthropic's controlled cybersecurity initiative that gives select organizations access to the unreleased model.

The catch: the Pentagon is doing this while actively working to complete a transition away from Anthropic entirely.

The Too-Dangerous-to-Release Model the Government Needs

Claude Mythos Preview is the model Anthropic described as so capable it posed a danger to the world. When Anthropic announced it on April 7, the company said it had decided "not to make it generally available" and would instead use it exclusively for defensive cybersecurity through Project Glasswing. The model can reportedly identify decades-old vulnerabilities across sprawling digital infrastructure that no human team or previous AI system could find.

Michael described the arrangement as a "temporary advantage" for the US government, acknowledging that competing models from OpenAI, xAI, and Google would soon match or exceed Mythos's cybersecurity capabilities. OpenAI announced its own competing initiative, Daybreak, just one day before the Pentagon's Mythos deployment was confirmed.

The Paradox of Essential and Unwanted

The Pentagon's relationship with Anthropic has been deteriorating for months. Anthropic was blacklisted from new government contracts earlier this year amid concerns about supply-chain risk and the company's governance structure. Two active lawsuits are contesting that decision, with Anthropic fighting to reverse the blacklisting through legal channels.

Yet here is the Department of Defense, using Anthropic's most advanced model to patch vulnerabilities in the very systems that protect national security. The logic is straightforward if uncomfortable: when you have access to the most capable cybersecurity tool in existence, you use it, even if you are planning to stop doing business with the company that made it.

Michael framed the risk in both directions. The increased capability of AI tools like Mythos means vulnerabilities can be addressed faster than ever. But those same capabilities mean adversaries with access to similar models can exploit vulnerabilities just as quickly. The window between discovery and patch, once measured in weeks or months, is now measured in hours.

A Cybersecurity Arms Race With No Offramp

This is not just a Pentagon problem. Banks are scrambling to patch thousands of vulnerabilities that Mythos has already found in their systems through separate Glasswing partnerships. The broader pattern is the same everywhere Mythos touches: organizations discovering that their digital infrastructure has far more holes than they knew, with far less time to fix them than they expected.

The Pentagon's decision to deploy the tool it is trying to walk away from tells you everything about the current state of AI cybersecurity: the technology is moving faster than the procurement bureaucracy, faster than the legal disputes, and faster than any institution's ability to form a coherent long-term vendor strategy.

For now, the US government is patching its networks with Anthropic's AI while its lawyers argue in court over whether to keep doing business with Anthropic at all. If that sounds like a contradiction, it is because the alternative, leaving known vulnerabilities unpatched while the paperwork sorts itself out, is worse.