The Trump White House is drafting an executive order that would bring Anthropic's Mythos model back into civilian federal agencies despite the Pentagon's "supply chain risk" classification of the company. The same week, Anthropic asked to expand Mythos private-sector access from approximately 50 firms to approximately 120; the White House blocked that on compute-strain grounds while drafting paperwork to expand its own access. The NSA is already using Mythos for cyber-zero-day work despite the DoD blacklist. [1]

The frame here is not turf. The federal government is producing one set of paperwork that designates Anthropic a supply-chain risk, a second set that quietly authorizes NSA usage despite the designation, and a third set that prepares to readmit the company to civilian agencies through executive action. Three contradictory paperwork streams from the same federal apparatus is what dependency looks like when it cannot be admitted publicly. The simpler explanation is that the U.S. government cannot find an alternative to Mythos for the cyber-vulnerability work it has been routing through Anthropic, so the "supply chain risk" classification was always a negotiating posture. The governance memo missing from this paper's Apr 30 dormancy frame now resolves cleanly: there is no governance memo because there is no alternative. [2]

The Pentagon's own posture sits inside this contradiction. The Department of War — Hegseth's renamed Pentagon — has spent the past several months suing Anthropic and being sued by Anthropic in federal court over the supply-chain-risk designation. A federal judge in San Francisco blocked the Pentagon order in late 2025 on the grounds that officials had likely violated the law and retaliated against Anthropic for speaking publicly about how it wanted its technology used. The litigation has produced a near-permanent legal-procedural standoff between the Pentagon and the company. The NSA, which sits inside the DoD reporting structure, has nonetheless been deploying Mythos for software-vulnerability discovery against Microsoft and other widely used commercial software — work the federal government considers strategically essential and cannot delegate to OpenAI, Google, or any other current foundation-model lab. [3]

The proposition the executive-order paperwork tests is whether the White House can route around its own Pentagon's stated security position. The mechanism is the use of presidential authority to direct civilian agency procurement — Treasury, State, Commerce, NSC — independent of DoD's Foreign Risk classification authority. If the EO is signed, the contradictory paperwork becomes a binding artifact: the United States government formally declares Anthropic a supply-chain risk inside DoD while contracting with the company through every other agency. That construction is unstable, but it is also the construction the White House is now drafting toward.

The compute-strain rationale for blocking Anthropic's private-sector expansion is the more honest piece of the framing. Mythos is reportedly capacity-constrained — Anthropic has described shipping non-Mythos products (Sydney/NEC/Amazon partnerships, the 5GW Glasswing data center, financial-services deployments) while keeping Mythos behind a tight private-sector access list. The federal government wanting more of that constrained capacity is the same demand-side pressure the private sector is asking for. The fact that the White House is preparing to expand its own access while denying Anthropic's request to expand third-party access tells you that compute-allocation is now a federal-versus-private contest, not a federal-versus-other-federal one. [4]

For Anthropic, the asymmetric frame is that the company is the only foundation-model lab that has produced a model the federal government considers irreplaceable for a specific national-security mission. OpenAI has the larger commercial footprint and is now in product-liability litigation in California and a criminal probe in Florida; Google has Gemini's commercial deployment but no comparable national-security-mission claim; xAI has Grok inside the federal procurement process but has not claimed cyber-zero-day capability the way Anthropic's NSA work suggests Mythos has. The Pentagon's "supply chain risk" classification was the negotiating leverage for a counterparty that knew the federal government could not actually replace it.

The executive-order draft, if signed, would convert that leverage into an explicit single-buyer dependency. The federal government would hold Anthropic in a procurement structure that binds the company to specific use-case constraints (no autonomous weapons deployment, no domestic surveillance routing) while paying for capacity the company says cannot be expanded for the broader market. That is a sole-source procurement relationship, dressed up as a national-security carve-out. Sole-source relationships in federal procurement are politically fragile and structurally durable — they survive everything except a credible alternative.

Whether the EO arrives this week, next month, or never is now the watch item. The fact that the paperwork is being drafted is the news. The Pentagon's continuing legal posture is the contradicting news. The NSA's quiet usage is the resolving news.

-- ANNA WEBER, Berlin