Detailed Analysis
The National Security Agency has quietly been granted access to Anthropic's Mythos Preview model — the company's most advanced and cybersecurity-focused AI system — even as the broader Department of Defense has designated Anthropic a "supply-chain risk" and is actively challenging the company in court. Mythos Preview, which Anthropic restricted to approximately 40 organizations and publicly named only a dozen of, is being used by the NSA primarily to scan digital environments for exploitable vulnerabilities. The UK's AI Security Institute has also confirmed access to the model, suggesting a degree of allied intelligence-community coordination around the technology. The revelation, reported on April 19, 2026, exposes a striking inconsistency within the U.S. government's posture toward Anthropic: different arms of the national security apparatus are simultaneously treating the company as both a threat and an indispensable asset.
The contradiction at the heart of this situation stems from a dispute that originated in February 2026, when the Pentagon demanded unrestricted access to Claude for military applications. Anthropic refused, citing specific concerns about enabling mass domestic surveillance and the development of autonomous weapons systems. That refusal led the Department of Defense to label Anthropic a supply-chain risk — a designation that carries legal and procurement implications — and to pursue litigation. Yet the NSA's quiet adoption of Mythos Preview suggests that operational demand for frontier AI capabilities within the intelligence community is outpacing the Pentagon's formal policy stance, creating a fragmented and potentially contradictory whole-of-government approach to AI procurement and risk assessment.
Diplomatically, the situation appears to be in flux. Anthropic CEO Dario Amodei recently met with White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent, with the White House characterizing the meeting as productive. This signals that Anthropic's relationship with the Trump administration may be warming even as the Pentagon conflict remains unresolved. The meeting is notable because it elevates the Anthropic dispute from a procurement-level disagreement to a matter of direct White House engagement, suggesting that senior administration officials view the company's technology — and its cooperation — as strategically significant enough to warrant executive-level diplomacy.
Mythos Preview's unique capabilities help explain why demand for it persists despite the political friction. The model's advanced coding ability and capacity for autonomous task execution give it an unprecedented aptitude for identifying software vulnerabilities and devising methods of exploitation — precisely the kind of capability that is valuable to signals intelligence organizations like the NSA. However, those same attributes have also generated substantial concern among cybersecurity researchers and policymakers, as the model could theoretically lower the barrier for offensive cyber operations if access were to be misused or inadvertently broadened. Anthropic's decision to limit distribution to a curated set of roughly 40 organizations reflects an attempt to maintain oversight, though the inclusion of a major intelligence agency in that group will inevitably intensify scrutiny of whether such controls are sufficient.
The broader trend this episode illustrates is the growing difficulty AI developers face in managing relationships with state actors at a moment when frontier models are increasingly mission-critical to national security. Anthropic finds itself in an unusual position — locked in a legal dispute with one branch of the U.S. military while apparently deepening ties with another agency and engaging directly with the White House. This fragmentation mirrors a wider pattern in the AI industry, where companies must navigate conflicting demands from governments that simultaneously want to harness AI's power, restrict adversaries' access to it, and impose usage conditions that developers may find ethically or commercially untenable. How Anthropic resolves its Pentagon standoff while sustaining relationships with other national security customers is likely to set an important precedent for how AI companies manage state-level partnerships going forward.
Read original article →