NSA Runs Anthropic Mythos AI Despite Pentagon Blacklist

The Trump administration branded Anthropic a "national security menace." The Pentagon tried to cut the company off from federal contracts entirely. Yet inside the National Security Agency (NSA) — America's primary signals intelligence organization (the agency responsible for intercepting foreign communications, breaking enemy codes, and monitoring adversary networks) — engineers are actively deploying Anthropic's most powerful AI model: Mythos Preview.

This contradiction, first reported by Axios on April 20, 2026, exposes a sharp fault line running through the U.S. government's approach to AI: political caution at the executive level, technical urgency at the operational level. It's unfolding as Anthropic CEO Dario Amodei runs a quiet diplomatic offensive — White House meetings with Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent, both sides describing the exchange as "productive and constructive."

A U.S. Government Divided on Anthropic AI Policy

The Pentagon's attempted blacklist of Anthropic didn't emerge in a vacuum. Earlier in the Trump administration, officials flagged Anthropic as a potential national security risk — language that, under standard DoD (Department of Defense — the U.S. military's executive body overseeing all federal defense procurement) rules, triggers formal vendor exclusion processes. Anthropic apparently landed squarely on that list.

But the NSA operates with significant technical autonomy. Its mission demands the best available tools, political considerations aside. When Mythos Preview became available, NSA technical teams apparently concluded the capability gap was too large to sacrifice on political grounds alone.

• Pentagon/DoD position: Blacklist Anthropic over national security and political concerns
• NSA position: Actively deploy Mythos Preview for operational intelligence work
• The result: 2 agencies under the same national security umbrella running opposite Anthropic policies simultaneously

This kind of internal contradiction is rare in federal procurement. Most agencies follow centralized guidance from the OMB (Office of Management and Budget — the executive branch office that sets and oversees federal spending policy). For the NSA to diverge from Pentagon direction openly suggests either a deliberate institutional workaround, or a genuine belief among NSA leadership that Mythos delivers capabilities no politically acceptable alternative can match right now.

What Anthropic Mythos Preview Is — And Why the NSA Wants It

Anthropic describes Mythos as its "powerful new AI model" — positioned at the frontier tier (the cutting edge of AI, above standard commercial offerings like Claude Sonnet which most developers use today). Axios confirmed NSA is using it but did not disclose deployment specifics: scale, specific applications, integration depth, or whether this is a limited pilot (a small-scale test with restricted scope) or a full production rollout.

For intelligence contexts, frontier models (AI systems operating at the absolute leading edge of current technology) are valued for capabilities standard models can't reliably deliver:

• Deep reasoning chains: Multi-step analytical work — critical for threat assessment, signals interpretation, and attribution analysis
• Long context windows: Processing enormous document sets — tens of thousands of intercepted communications — without losing analytical coherence across the full dataset
• Adversarial robustness: Accuracy on tasks specifically designed to mislead or fool the AI — a baseline requirement in any intelligence environment
• Multilingual processing: Handling foreign-language material at scale without requiring a human translator in the loop for each document

The "Preview" designation is also revealing: it suggests Mythos is still in limited availability, meaning the NSA may hold early-access privileges that standard commercial customers don't have. That kind of privileged vendor relationship makes any formal blacklist dramatically harder to enforce — you can't easily cut off a company when your own agencies are embedded in their early-access partner network.

Dario Amodei's White House Play

While the NSA-Pentagon conflict plays out inside government bureaucracy, CEO Dario Amodei has been working the political angle directly. His White House visit — meetings with Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent, 2 of the most influential officials in the current administration — is a classic Washington move: when procurement threats materialize at the agency level, escalate directly to the executive office.

The official framing of "productive and constructive" from both sides is diplomatic détente language (careful signaling of de-escalation without committing to specific policy reversals). Given that Anthropic entered those meetings tagged as a "national security menace," even détente language represents real movement. The White House isn't using that language with vendors it plans to keep on the blacklist.

The arguments Amodei almost certainly made in that room:

• Operational reality: NSA is already running Mythos. The contradiction exists on record whether anyone formally acknowledges it or not
• Competitive stakes: Blocking U.S. AI companies from government work advantages Chinese AI development — Baidu, DeepSeek, and state-backed models face no such restrictions inside their own government
• Economic weight: At an $800 billion valuation with active investor interest, Anthropic represents enormous U.S. economic value; treating it as a threat carries direct strategic costs
• The precedent problem: If political characterizations routinely override NSA technical judgment, it undermines the agency's ability to maintain intelligence superiority going forward

The $800 Billion Anthropic Valuation Backdrop

The timing of investor interest at an $800 billion valuation — surfacing alongside this story — isn't incidental context. U.S. government AI contracts represent tens of billions in annual procurement across defense, intelligence, and civilian agencies. Anthropic's ability to access that market directly affects how institutional investors (large funds and financial institutions deploying capital at scale, looking for long-term growth ceilings) price the company.

A sustained DoD blacklist spreading to other agencies would cap Anthropic's addressable market significantly. But if Amodei's White House diplomacy produces formal policy reversal — and NSA's active adoption creates a de facto precedent — the blacklist dissolves, and those growth projections expand accordingly. Investors pricing in $800 billion appear to be betting political friction is temporary and technical capability wins long-term. The NSA's decision to run Mythos now makes that bet look considerably less speculative.

What the NSA-Pentagon Split Signals for Enterprise AI Automation

Anthropic's situation is instructive beyond the company itself. The U.S. government is not a monolithic AI buyer. Different agencies have different risk tolerances, different technical requirements, and different relationships with the political leadership of any given administration. A DoD blacklist doesn't automatically bind the NSA, the CIA (Central Intelligence Agency), or the broader intelligence community's own procurement channels.

The broader pattern: when a tool's operational advantage is large enough, technical teams inside institutions will find ways to deploy it even against formal restrictions. The NSA just demonstrated that at the highest-stakes level possible — and the story broke publicly, which means the informal adoption is now a matter of documented record that future policy arguments will have to account for.

If you're tracking Anthropic's platform for reliability signals or evaluating AI automation tools for serious enterprise workloads, you can explore our guide to AI adoption for professional use cases. And watch the White House outcome closely: if Amodei's meetings produce concrete policy reversal — an official lifting of the "national security menace" label, or a DoD carve-out for intelligence community use — it would mark one of the fastest political rehabilitations in recent technology history. Mythos is not vaporware. It's running right now, inside one of the world's most demanding intelligence operations, while the politics around it remain officially unresolved.

Related Content — Get Started | Guides | More News