← Back to reports

The Algorithm That Refused

How a Contract Clause Triggered a Structural Break Between Anthropic and the Pentagon

Based on reporting from Axios, NPR, CNBC, and Washington Post as of February 28, 2026. Some operational details remain unconfirmed by official sources

On February 27, 2026, the Trump administration designated Anthropic a federal supply chain risk, effectively excluding the company from U.S. government work.

The trigger was not performance failure. It was a contractual disagreement over two restrictions: prohibitions on large-scale domestic surveillance and on fully autonomous lethal use without human oversight.

This is not primarily a story about AI ethics. It is a story about authority — specifically, who retains final control over military applications of frontier AI systems.

The Structural Fault Line

In July 2025, the Pentagon awarded AI contracts worth approximately $200 million to four companies: Anthropic, OpenAI, Google, and xAI.

Anthropic’s agreement included two explicit restrictions: its model could not be used for large-scale surveillance of U.S. citizens, nor for autonomous weapons without meaningful human oversight.

The other contracts did not include the same explicit company-level constraints.

That difference remained dormant until early 2026.

Claude, Anthropic’s model, was deployed within classified defense environments and, according to reporting by major outlets, was used in analytical support during the January 3 operation that resulted in the capture of Nicolás Maduro. Public sources describe its role as rapid processing of intelligence, intercepts, and operational data.

The model was considered operationally effective.

The dispute was not about performance. It was about scope.

The Ultimatum

In February, Defense Secretary Pete Hegseth reportedly informed Anthropic that continued participation would require removal of company-imposed restrictions in favor of a standard allowing “all lawful uses.”

Anthropic refused.

The company’s position was that its red lines had not interfered with any mission to date. The Pentagon’s position was that the existence of a private veto over potential military applications was unacceptable in principle.

From the Department’s perspective, uncertainty of control over a critical operational system is itself a liability.

This is where the conflict moved from operational to constitutional in character.

The Jurisdiction Question

The key distinction between Anthropic and OpenAI was not necessarily ethical stance, but locus of authority.

OpenAI publicly maintains similar guardrails regarding mass surveillance and autonomous lethal systems. However, its agreement with the Pentagon defers the definition of “lawful use” to government interpretation rather than reserving an explicit corporate override.

Anthropic attempted to retain a contractual mechanism that could restrict certain uses even if deemed lawful by the government.

The Pentagon interpreted that as a structural risk.

One side frames this as an ethical safeguard.
The other frames it as a private entity reserving potential veto power over sovereign military decisions.

Both interpretations are internally coherent.

The Musk Factor

xAI had already accepted the “all lawful use” standard.

This matters because of vertical integration. Elon Musk now controls:

Historically, military suppliers have been segmented across domains. Consolidation across launch, communications, and AI analysis represents a new configuration of private-sector strategic leverage.

This is not about ideology. It is about concentration of infrastructure.

What Anthropic Actually Did

Anthropic refused to remove two lines from a contract.

Whether one views that as principled restraint or strategic miscalculation depends on perspective. But it marks the first high-profile case of a major AI company declining expanded military latitude after operational deployment.

Dario Amodei has previously articulated concerns about concentrated lethal authority enabled by advanced AI systems — specifically the risk of reducing the number of human decision points in kinetic operations.

That concern is not abstract. It reflects a long-standing debate in military theory about automation and accountability.

The Strategic Implication

The deeper issue is not whether Claude was used effectively in Venezuela. It is whether the state will tolerate operational dependence on a system that a private board could constrain.

Historically, when a technology becomes central to national security, the state either absorbs it, regulates it tightly, or restructures the market to eliminate ambiguity of control.

From that perspective, the Pentagon’s reaction is not anomalous. It is predictable.

Anthropic lost the contract.

But the larger question remains unresolved:

Can a private AI company retain enforceable limits once its systems become embedded in national defense infrastructure?

The answer to that question will shape the governance architecture of military AI more than any single operation.