The Pentagon’s standoff with Anthropic is not just a procurement dispute. It is a preview of the next phase of the AI era, when the state will treat frontier models like critical infrastructure and companies will treat their safety policies like constitutional rights. The conflict forces an uncomfortable question: when a model can scale power, who gets to set the red lines, and what happens when the government decides your red lines are a national-security vulnerability?

Over the past week, reporting described an escalating confrontation between the U.S. Department of Defense and Anthropic about whether Claude should be available for “all lawful purposes,” versus whether Anthropic could hold firm on two specific restrictions: no use for domestic mass surveillance and no use in fully autonomous weapons. NPR, CNN, and others described a deadline-driven negotiation, followed by moves to cut off use of Anthropic systems and to frame the company as a “supply chain risk.”[1][2]

Anthropic, for its part, published a statement arguing that current frontier models are not reliable enough for fully autonomous weapons, and that mass domestic surveillance violates fundamental rights.[3] Major outlets reported that the executive branch directed agencies to stop using Anthropic, while the Defense Department leaned on a “supply chain risk” framing that would ripple across contractors.[4][5]

It is tempting to read this as partisan theater or Silicon Valley melodrama. But the underlying tension is structural. Frontier AI is becoming a general-purpose capability. General-purpose capabilities always get absorbed into the machinery of national power: first as tools, then as dependencies, then as choke points.

“All lawful purposes” sounds bounded, but it is effectively unbounded because law is a moving target.

If you ship a model under “all lawful purposes,” you are not just agreeing to today’s use cases. You are agreeing to the future interpretations of lawful. You are also agreeing to the enforcement incentives of the purchaser.

This is why companies create usage policies. It is not to virtue signal. It is to define a minimal set of constraints that survive political cycles.

For a frontier model company, the usage policy does three jobs:

In the industrial era, your constitution was your cap table and your bylaws. In the AI era, your constitution is your deployment policy.

When a government asks you to drop the constitution, it is not negotiating a clause. It is asking you to change your identity.

From the Pentagon’s perspective, a supplier that reserves the right to say no is a supplier that can fail at the worst time.

Defense systems are built around optionality:

A frontier model that is “too powerful to ignore” but “too principled to deploy freely” looks, to the state, like a strategic dependency that can be weaponized against the state.

So the state will respond the only way it knows: convert it into something legible.

That means compliance language, leverage, and procurement tools.

The “supply chain risk” label functions like an immune response. It tells the broader system: route around this dependency.

Whether or not you agree with any particular foreign policy, the pattern is consistent across eras:

The “current war in Iran” is one example of the kind of pressure environment that compresses decision cycles and raises the demand for intelligence, surveillance, and targeting capacity.

In such contexts, the argument “we will not do mass surveillance” collides with the reality that every institution in a high-threat environment will try to expand its sensing and prediction capabilities.

If frontier AI helps do that, it will be pulled into that orbit.

Models will commoditize. Governance will differentiate.

Two companies can have similar benchmarks but very different answers to:

The procurement dispute is actually a dispute about who owns the governance layer.

Both sides believe the other side is asking for an unacceptable delegation of power.

Even if current leadership offers assurances, the contract language matters because leadership changes.

The moral hazard is simple: once the tool exists and the pathway is open, incentives will exploit it.

Also, bureaucracies do not behave like individuals. They behave like systems:

A supplier that relies on verbal assurances is not doing governance. It is doing hope.

This is the strongest argument.

A world in which democracies self-restrict while authoritarian regimes deploy freely is a dangerous asymmetry.

But the answer is not necessarily “drop all restrictions.” It is “build restrictions that scale.” The hard work is designing usage constraints that preserve deterrence and defense while preventing abuses.

The alternative is a race to the bottom in which every capability becomes permissible because someone else might do it.

That logic does not produce security. It produces inevitability.

B Hari

Simplicity with substance
www.bhari.com