NSA Uses Mythos Preview as Pentagon Labels Anthropic a 'Supply-Chain Risk'

Key Points

• The NSA is actively using Claude Mythos Preview for defensive cyber operations
• The Pentagon has simultaneously designated Anthropic as a "supply-chain risk"
• White House officials have met with Anthropic leadership about safe government use
• The OMB is developing safeguards for broader government access
• The contradiction highlights unresolved tensions around AI in national security

The Contradiction at a Glance

In a development that underscores the complexity of AI governance, two branches of the U.S. government hold opposing stances on Claude Mythos Preview. The National Security Agency uses it actively for cyber defense. The Department of Defense has formally labeled the company that made it a security risk.

This isn't a bureaucratic oversight — it reflects genuine disagreement about how to handle AI systems that are simultaneously powerful tools for national defense and potential vectors for new risks.

Pentagon's "Supply-Chain Risk" Designation

The Department of Defense formally designated Anthropic as a "supply-chain risk" in mid-April 2026, following what reports describe as a contract dispute. The Pentagon's specific concerns reportedly include:

• The potential for the model to be used in autonomous weapons development
• Risks related to mass surveillance applications
• Concerns about foreign government access through Anthropic's cloud partners
• The precedent of a private company controlling access to strategically significant AI capabilities

The supply-chain risk label, while not an outright ban, complicates the ability of DoD agencies to procure Anthropic's services and creates additional compliance requirements.

NSA's Defensive Use

Despite the Pentagon's position, reporting from multiple outlets confirms that the NSA has been actively using Claude Mythos Preview through what appears to be a separate authorization pathway. The agency's use case is specifically defensive: using the model's exceptional vulnerability detection capabilities to identify weaknesses in critical national infrastructure before adversaries can exploit them.

According to reports, the NSA views Mythos Preview as a uniquely valuable tool for cyber defense precisely because of the capabilities that concern the Pentagon. The model's ability to discover zero-day vulnerabilities at scale — including the 27-year-old OpenBSD bug — makes it potentially more effective than any human team at proactive infrastructure protection.

White House Engagement

The tension between agency positions has prompted White House involvement. Reports indicate that Anthropic's leadership — including CEO Dario Amodei — has engaged in discussions with:

• The White House Chief of Staff regarding policy frameworks for AI in government
• The Treasury Secretary regarding economic and financial implications
• OMB officials developing safeguards for authorized government use

The Office of Management and Budget is reportedly working on a framework that would allow authorized government agencies to access a modified version of Mythos Preview with appropriate controls, potentially resolving the NSA-Pentagon standoff.

Broader Implications

This situation raises several questions that extend beyond a single model:

Who decides access? When an AI system has significant national security implications, the question of who controls access — a private company, the military, intelligence agencies, elected officials — becomes urgent and unresolved.

Dual-use dilemma: Mythos Preview's cybersecurity capabilities are inherently dual-use. The same ability to find vulnerabilities defensively could theoretically be redirected offensively. This makes traditional arms-control frameworks difficult to apply.

Precedent-setting: How the U.S. government resolves this particular tension will likely influence how other governments handle similar situations with their own frontier AI providers.

We will update this article as the situation develops. For background on the model's capabilities, see our system card analysis and explainer page.