Detailed Analysis
The National Security Agency is reportedly deploying Anthropic's Mythos Preview model for cybersecurity operations including vulnerability scanning, a development that stands in sharp tension with the Department of Defense's formal designation of Anthropic as a "supply-chain risk." Announced by Anthropic in early April 2026, Mythos was described as the company's most capable model to date for coding and agentic tasks, with particular strength in cybersecurity applications. Citing the model's potential for offensive cyberattack capabilities, Anthropic withheld it from general public release, restricting access to approximately 40 organizations, of which only 12 have been publicly named. The NSA is among the undisclosed recipients, and the UK's AI Safety Institute has separately confirmed its own access to the model. Neither Anthropic, the NSA, nor the DoD offered any comment on the reported arrangement.
The NSA's use of Mythos exists in direct contradiction to the Pentagon's broader campaign against Anthropic, which began crystallizing in early 2026 following a breakdown in contract negotiations. The DoD had sought unrestricted access to Anthropic's Claude models for applications including surveillance and autonomous weapons systems — demands Anthropic refused on grounds aligned with its stated safety principles. The Pentagon subsequently began severing ties with Anthropic and pressuring its own suppliers to do the same, while simultaneously advancing legal arguments that Anthropic's tools represent a threat to national security. The fact that the NSA — itself a major component of the U.S. intelligence and defense apparatus — is actively using the very technology the Pentagon is attempting to restrict reveals a significant lack of coordination within the national security establishment and suggests that operational cybersecurity imperatives are, at least in some corridors, overriding institutional policy.
The episode also coincides with signals of a potential diplomatic thaw between Anthropic and the Trump administration more broadly. Anthropic CEO Dario Amodei met last Friday with White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent, with the discussion described as productive. This meeting suggests that while Anthropic's relationship with the DoD remains adversarial, the company is actively cultivating separate channels of influence at the highest levels of the executive branch. The parallel tracks — Pentagon blacklist on one side, White House engagement and NSA adoption on the other — reflect the fragmented and frequently inconsistent nature of U.S. government AI policy in 2026.
The broader context is one in which frontier AI models with specialized capabilities are rapidly becoming embedded in national security infrastructure, often faster than governing frameworks can accommodate. Anthropic's deliberate decision to limit Mythos distribution rather than release it publicly represents an attempt to exercise supply-side control over dual-use AI — a strategy that simultaneously positions the company as a responsible actor while granting it significant leverage over which institutions gain access to its most powerful tools. The reported surge in vulnerability disclosures in early 2026, which Hacker News communities have linked to AI-driven security research and specifically to Mythos-class capabilities, underscores how profoundly such models are reshaping the offensive and defensive cybersecurity landscape. The NSA's quiet adoption, even amid official hostility from the Pentagon, is a strong indicator that governments worldwide will continue reaching for the most capable AI tools available regardless of the institutional friction involved.
Read original article →