Intelligence Agencies Bypass Defense Procurement Hurdles to Access Restricted AI Models

What Happened

The National Security Agency (NSA) is reportedly utilizing Anthropic’s “Mythos” AI model despite the Pentagon formally designating the firm as a “supply chain risk.” Mythos, currently restricted to approximately 40 organizations due to its advanced offensive coding and agentic capabilities, is being deployed for intelligence operations. This usage persists while Anthropic remains engaged in litigation against the Department of Defense regarding forced access to proprietary training data and source code.

Why It Matters

First-order: The intelligence community is prioritizing operational superiority over standardized Pentagon procurement compliance. The NSA is essentially bypassing the “risk” designation to gain access to superior autonomous capabilities that off-the-shelf defense models lack.

Second-order: This creates a fragmented federal procurement environment. Companies that prioritize security and intellectual property protections—such as Anthropic—may find themselves in a “gray market” relationship with specific government agencies, while traditional defense contractors continue to lobby for restrictive compliance frameworks that favor legacy incumbents.

Third-order: We are seeing the decoupling of national security priorities from defense procurement bureaucracy. Intelligence agencies are shifting toward a ‘buy best-in-class’ strategy, signaling that vendors who refuse to open their black boxes may still win high-value government contracts if their model performance remains unmatched.

What To Watch

• Precedent for Proprietary Protections: Whether Anthropic’s legal strategy against the DOD successfully creates a framework for AI vendors to shield source code and training data from mandatory government audit.
• Fragmented Federal Procurement: How the Department of Defense addresses the defiance of its risk designations by its own subsidiary agency, the NSA.
• Expanded Access: Potential updates to the Mythos restricted access list as the model moves out of the preview phase and into production-grade intelligence workflows.