It is Day 5 of what court filings have made into a documented contradiction. The Pentagon's current AI roster — OpenAI, Google, Microsoft, Amazon, Nvidia, SpaceX and Reflection — still excludes Anthropic, citing a "supply chain risk" designation that the department has argued in court. The National Security Agency, according to Anthropic court filings now in the record, has been using Mythos Preview, Anthropic's frontier model, on the same calendar inside which the Pentagon's exclusion stands. [1][2][3]

The paper's May 10 feature on Anthropic's loss of the Pentagon contract and the civilian-EO workaround framed this as two-track government in writing. The cycle has now run five days without a White House executive order on civilian agencies, and the trip toward the Trump-Xi summit on May 14-15 is three days out with the Iran agenda taking precedence. Axios reported the EO is being drafted at the White House to permit civilian-agency Anthropic use around the Pentagon's supply-chain designation; Nextgov confirmed the EO mechanism is the working approach. [1][2]

Mario Nawfal's post compressing the contradiction — the NSA using Mythos Preview while the Pentagon argues in court that Anthropic is a supply chain risk — is the position the paper finds difficult to dispose of as either a misreading or a routine bureaucratic mismatch. Two agencies of the same government cannot simultaneously hold both positions in writing without one of them being wrong, or both being right inside a system that has no single point of decision. Reuters Legal had already reported that Anthropic was in discussions about Mythos with the administration after the Pentagon's cut-off. The administration was not a single party then. It is less of one now.

The Pentagon's argument, in Defense News's account, was procurement-procedural. The supply-chain risk designation emerged from a contract dispute; the technical merits of Mythos are not what the department challenged. The department challenged whether Anthropic's structure — its safety-focused architecture, its public-benefit corporate form, its acceptable-use policies that constrain certain government work — produced a vendor that could meet defense supply-chain obligations. The designation is administrative law. The exclusion is operational. Both can be defended on bureaucratic terms. [3]

What cannot be defended on bureaucratic terms is the simultaneity. The NSA, organizationally inside the Department of Defense, is using a vendor the Department of Defense's procurement counsel has named a supply-chain risk in court. The White House civilian EO would route around this by carving out civilian agencies. The EO has not been signed at Day 5. Bessent and Wiles met with Anthropic CEO Dario Amodei in late April; "table reads" of draft language continue this week. The mechanism exists in draft. The mechanism does not exist in the Federal Register.

Inside Hannah Arendt's vocabulary, what is being demonstrated is the difference between bureaucratic rationality and political accountability. The Pentagon's exclusion has rational procurement grounds. The NSA's use has rational operational grounds. The civilian EO has rational political grounds. Together, the three produce a regime in which any external actor — an adversary, a customer, a journalist — encounters an inconsistent posture because the system has no actor who is responsible for the consistency. The contradiction is not a glitch. It is what the institutional architecture produces when reverence for individual-agency autonomy outruns the apparatus that would have to harmonize the agencies.

This matters because the Trump-Xi summit on May 14-15 will absorb the question whether U.S. AI exports remain consolidated under one regime or partition into civilian and military tiers with separate vendor lists. Chinese counterparts cannot read U.S. AI policy through a single authority because the authority is itself two-track. The summit's Iran agenda, named by Treasury Secretary Bessent over the weekend, is the dominant agenda. The AI-vendor question will sit on the agenda underneath it. Whether the civilian EO is signed before the summit will determine whether U.S. negotiators arrive in Beijing with a coherent AI position or with two.

Anthropic's competitive position acquires its own asymmetry from the regime. The company is excluded from a procurement track its competitors occupy, which is a commercial disadvantage. The company's safety architecture is being cited as the reason for the exclusion, which is a regulatory liability rather than a commercial one. The same architecture is being used by the NSA because it is operationally useful. The company sits inside a structure where the same property — Anthropic's chosen corporate form and safety policies — produces both the loss of the contract and the gain of the customer.

Day 5 ends with neither a signed EO nor a Pentagon reversal. The court filings stand. The NSA usage stands. The summit clock runs.

-- ANNA WEBER, Berlin