The NSA is reportedly using Anthropic's new model Mythos

The NSA is reportedly using Anthropic's new Mythos Preview model, even as Anthropic remains in a legal and policy dispute with the Pentagon and Trump administration. Anthropic says Mythos is a general-purpose model that is "strikingly capable at computer security tasks," and the company has given access to roughly 40 organizations. The news follows CEO Dario Amodei's meeting with White House officials to discuss Mythos, while court battles over the government's "supply chain risk" designation continue.

Analysis

This looks less like a clean commercial win for the model vendor and more like evidence that the government is fragmenting its AI procurement stance by mission set. The near-term beneficiary is not necessarily the company’s equity directly, but the broader category: security-sensitive agencies can justify selective adoption of frontier models when the use case is bounded to cyber defense, even while policy fights continue elsewhere. That creates a two-tier market in enterprise AI—general-purpose workloads remain exposed to political risk, while cyber, code, and controlled-inference workloads become the first durable government budget line.

The second-order effect is competitive: vendors with stronger model performance in security tasks and better compliance wrappers gain share from horizontally positioned peers that cannot clear the same trust bar. If this pathway broadens, the real winners are the infrastructure layer and security integrators that sit between the model and the agency—systems integrators, cloud hosts, and data-governance tooling—because agencies will prefer to buy “guardrailed access” rather than raw model subscriptions. The legal overhang, however, means adoption can rise even as procurement headlines remain toxic, so market reaction in the name itself may lag actual usage.

Catalyst-wise, the next 2-6 weeks matter more than the next quarter: any further White House engagement or additional agency adoption would validate a controlled re-opening, while a fresh injunction or renewed Pentagon pushback would likely re-freeze deal flow. The tail risk is reputational and regulatory spillover—if one agency’s use is framed as an exception, competitors may demand identical treatment, increasing the odds of slower, more complex approvals across the sector. That argues for trading the policy dispersion rather than the headline itself.

Consensus is probably underestimating how quickly cyber-focused AI can move from pilot to sticky spend because the ROI is immediate and easier to defend than consumer or general enterprise deployments. What’s overdone is the idea that political friction blocks all revenue; what’s underdone is the risk that this becomes a precedent for compartmentalized adoption, benefiting a small set of vendors while widening the moat against everyone else.

Key Decisions for Investors

• Long security-vertical AI exposure vs broad AI beta for 1-3 months: buy a basket of cyber/software names with government penetration and short a diversified AI index proxy; thesis is that controlled federal adoption flows to compliance-heavy vendors faster than to generalist model names.
• If available in the portfolio universe, pair long AI infra/security enablers (cloud, data governance, SASE, endpoint/security orchestration) against a short in the most policy-sensitive frontier-model exposure; target is a 5-10% relative move on any additional agency adoption headline.
• For event-driven optionality, consider short-dated call spreads in leading cyber-defense software names ahead of the next White House or agency update; favorable if the market starts pricing a broader procurement thaw, with defined downside if the story remains isolated.
• Avoid chasing the headline in the directly mentioned model vendor until legal clarity improves; use any post-news weakness as a hedge opportunity rather than an outright long, since revenue visibility is improving slower than usage evidence.