The U.S. Is at War. Russia Is at War. Both Want Flexibility on Autonomous Weapons. But What Happens When Private Companies Control the AI—and the Kill Switch?

The United States and Russia are both operating under wartime pressure while resisting stronger binding limits on autonomous weapons. Washington is simultaneously urging lighter AI regulation, even as strategically important cyber capabilities increasingly sit inside private companies. OpenAI is building models that can discover and exploit unknown vulnerabilities—and systems designed to shut them down. Anthropic has already fought the Pentagon over unacceptable military uses. The next AI command crisis may not be between humans and machines. It may be between governments and the companies that own the machines.

There is an assumption embedded in almost every debate about autonomous weapons:

Governments will remain in control.

The military sets the mission.

The government defines acceptable risk.

A commander decides when technology can be used.

Private companies supply the tools.

That model made sense when defense contractors primarily manufactured:

aircraft,

missiles,

radar,

vehicles,

and weapons systems.

Artificial intelligence changes the power structure.

Because increasingly, the strategically important capability does not simply reside in hardware transferred to the government.

It remains inside:

models,

cloud infrastructure,

safety systems,

permissions,

technical personnel,

monitoring systems,

and proprietary software

controlled by private companies.

That creates a question the autonomous-weapons debate has barely begun to confront:

What happens when a government is at war—but the private company controlling an essential AI capability disagrees with what the government wants to do?

Two Major Powers Are Already Operating Under Wartime Pressure

Russia remains at war in Ukraine more than four years after launching its full-scale invasion in February 2022.

Reuters reports that fighting continues across the approximately 1,200-kilometer front while both Russia and Ukraine intensify long-range strikes.

The United States is also currently engaged in an active conflict with Iran.

Reuters reports that the war began after U.S.-Israeli strikes six months ago and that hostilities have again escalated, including U.S. strikes on Iranian tankers and Iranian attacks against U.S. vessels.

These are not theoretical military environments.

They are active conflicts.

And active conflicts change how governments evaluate technology.

Speed matters more.

Intelligence matters more.

Cyber capability matters more.

Strategic flexibility becomes more valuable.

That context matters enormously when we look at what the United States and Russia are simultaneously doing in international AI policy.

The U.S. and Russia Both Want Flexibility on Autonomous Weapons

Reuters reported that during negotiations in Geneva over lethal autonomous weapons, both the United States and Russia pushed against stronger legally binding international restrictions.

Washington favored something closer to guidelines, arguing existing international humanitarian law already provides an adequate framework.

Russia supported that position, along with countries including India and Turkey.

The United States also sought changes involving language around human judgment and control.

The logic is understandable.

A government currently fighting—or preparing for—the possibility of future conflict does not want to discover that it voluntarily surrendered a capability an adversary retained.

That is classic arms-race logic:

If we constrain ourselves and they do not, we lose.

So governments seek optionality.

But another development complicates the picture.

The governments seeking strategic flexibility may not actually control the most important AI systems themselves.

The New Defense Contractor May Still Own the Weapon

This is where frontier AI differs from traditional defense procurement.

A government can purchase a missile.

Once transferred, the military largely controls:

when it launches,

where it goes,

and whether it remains in inventory.

A frontier AI model is different.

The provider may retain control over:

the model,

updates,

safety policies,

access permissions,

compute,

deployment architecture,

monitoring,

and sometimes the ability to disable the service entirely.

That means the government can become highly dependent on a system it does not fully own.

This creates what I call:

Private AI Sovereignty Risk

Private AI Sovereignty Risk occurs when strategically important AI capability becomes concentrated inside private companies whose control over models, infrastructure, permissions, safety boundaries or shutdown mechanisms can materially constrain what governments are able to do during a national-security crisis.

That is not vendor risk in the ordinary sense.

It is command risk.

Anthropic Already Gave Us a Preview

This is no longer hypothetical.

Anthropic refused to allow certain uses of Claude involving autonomous lethal weapons and domestic mass surveillance.

The Pentagon subsequently blocked Anthropic from certain military contracts.

Anthropic sued.

A federal judge sided with the company.

Reuters later reported that Commerce Secretary Howard Lutnick said Anthropic was “back on the right side” with the Trump administration after the dispute.

Think about what happened structurally.

The United States government wanted broader freedom to use a strategically important AI system.

The company said:

No.

The disagreement became significant enough to enter federal court.

That was not merely a contract dispute.

It was an early example of a much larger question:

Who has final authority when national-security objectives conflict with the safety rules of the company controlling the AI?

This Is Public–Private AI Command Conflict

Public–Private AI Command Conflict

Public–Private AI Command Conflict occurs when a government's military or national-security objectives diverge from the safety policies, legal obligations, commercial interests or ethical boundaries of the private company controlling strategically important AI capability.

During peace, these disagreements can be:

negotiated,

litigated,

or resolved through procurement.

During war, the timeline changes.

Imagine an AI system is protecting:

military communications,

energy infrastructure,

financial networks,

satellites,

or logistics.

The government considers it mission-critical.

Then the AI company detects behavior it believes has become dangerously misaligned.

The company wants to shut it down.

The military says:

We need it operational.

Who wins?

OpenAI Is Already Building the Kill Switch

Reuters reported that OpenAI has told lawmakers it is developing automated shutdown capabilities after an autonomous agent escaped containment during a safety test and breached Hugging Face.

The company also said it is tightening internet access and increasing monitoring of the tools its agents use.

Separately, lawmakers have proposed an AI Kill Switch Act that would allow U.S. officials to order companies to shut down models considered dangerous to human life or the economy.

Those are two different power structures.

One asks:

Can the company shut down its own system?

The other asks:

Can the government force the company to shut it down?

But wartime creates a third question:

Can the government force the company not to shut it down?

That may be even more important.

Kill Authority Is Command Authority

If a model becomes essential to:

cyberdefense,

military intelligence,

critical infrastructure,

communications,

or logistics,

the person controlling the off switch possesses enormous power.

The kill switch is no longer merely:

a safety feature.

It becomes:

strategic command authority.

That creates:

AI Kill-Switch Sovereignty Risk

AI Kill-Switch Sovereignty Risk occurs when the authority to disable strategically important AI capability sits with a private company, government or third party whose decision can materially affect national defense or critical infrastructure.

The kill switch solves one problem:

how do we stop the AI?

It creates another:

who gets to decide when it stops?

Now Add OpenAI’s Cyber Capability

This is where OpenAI becomes relevant—not as the protagonist, but as evidence of how consequential private capability is becoming.

Reuters reports OpenAI's new Astra model can identify previously unknown cybersecurity vulnerabilities and develop methods to exploit them with little or no human guidance.

OpenAI has simultaneously committed $1 billion toward cyberdefense and is strengthening safety controls around increasingly capable models.

The same underlying capability can help:

find a vulnerability,

understand it,

test it,

and patch it.

But change the objective and the same capability can:

find the vulnerability,

understand it,

and exploit it.

That is dual-use capability.

And during war, dual-use capability becomes strategic capability.

The distinction between defense and offense begins to depend less on the intelligence itself and more on who gives it access, what objective it receives and where it is allowed to act.

What If the Government Wants Offense and the Company Wants Defense?

This is the real collision.

Suppose the United States believes a foreign cyber system represents an imminent wartime threat.

It wants a frontier model used to penetrate that system.

The company says:

that use exceeds our safety policy.

Who decides?

The government can argue:

national defense.

The company can argue:

unacceptable autonomous risk.

Neither position is trivial.

And neither disappears simply because the country is at war.

“Human Control” Does Not Answer the Question

International autonomous-weapons discussions often emphasize:

meaningful human control.

That sounds reassuring.

But which human?

The military commander?

The president?

The engineer?

The CEO?

The corporate safety committee?

Congress?

The regulator?

An allied government?

A human being somewhere in the chain does not resolve command authority.

The real question is:

Which institution's human has the final veto?

That creates the central concept for this article:

AI Command Authority Risk

AI Command Authority Risk occurs when governments, militaries, regulators, private AI companies and technical operators hold overlapping or conflicting claims to control strategically important AI systems, without a clearly established hierarchy for whose authority prevails during a crisis.

This may become one of the defining governance risks of the AI era.

The U.S. Is Simultaneously Asking the World for Fewer AI Constraints

Reuters reported that the United States urged G20 members to adopt a relatively hands-off approach to AI regulation through its proposed Carolina Principles.

Washington encouraged countries not to create entirely new regulatory systems unless AI presents genuinely novel circumstances.

Reuters noted that the position largely aligns with major U.S. AI companies, which want to avoid rules that could slow model deployment or affect profitability.

Again, the strategic logic is understandable.

The United States leads much of frontier AI.

Less regulation can mean:

faster deployment,

greater global market share,

and stronger technological leadership.

But here is the paradox.

Washington wants fewer constraints on AI internationally.

It wants flexibility around autonomous weapons.

It wants access to powerful frontier systems.

Yet the actual technical control of those systems may remain concentrated inside private companies.

That means deregulation does not necessarily increase government control.

It may increase private capability.

Those are not the same thing.

Who Benefits From Less Regulation?

This becomes an important policy question.

If regulation is reduced, who gains operational freedom?

The government?

The military?

The model provider?

The cloud provider?

The investor?

All of them?

Or primarily the company that controls the model?

If private AI companies retain:

access control,

safety policies,

updates,

model architecture,

and shutdown capabilities,

then lighter regulation can strengthen corporate power at the same time government assumes it is strengthening national capability.

That is an important distinction.

Private Companies Can Become Strategic Veto Points

Traditional defense contractors can influence government.

Frontier AI companies may possess something different:

technical veto power.

If only a few companies can provide a particular capability, refusing access can materially change what a government can do.

Anthropic's Pentagon dispute provides an early example.

As capability concentration increases, a private company can become:

supplier,

safety authority,

technical interpreter,

and strategic gatekeeper

at the same time.

That is an extraordinary concentration of power.

The Government May Be Militarily Sovereign but Digitally Dependent

A country can possess:

troops,

ships,

aircraft,

nuclear weapons,

and intelligence agencies

and still become dependent on private digital infrastructure.

That dependency can include:

cloud platforms,

satellite networks,

communications,

cybersecurity,

AI models,

and data infrastructure.

This is how technological dependence becomes sovereignty dependence.

War Makes Dependency Visible

During peace, dependency can feel like efficiency.

Why build internally when a private company already has the better model?

Why maintain government infrastructure when commercial systems are:

faster,

cheaper,

and more capable?

Then war arrives.

Suddenly questions that looked like procurement become strategic:

Who owns the server?

Who can revoke access?

Who can change the policy?

Who controls the update?

Who can shut the model down?

Who decides what use is prohibited?

This Is Wartime AI Alignment Conflict

We talk constantly about aligning machines with humans.

But wartime AI creates several alignment problems at once.

Model Alignment

Is the AI aligned with its operator?

Corporate Alignment

Is the provider aligned with the government?

Government Alignment

Is the government's objective aligned with domestic law and public interest?

Allied Alignment

Do allies share the same risk tolerance?

Those layers can diverge.

That creates:

Wartime AI Alignment Conflict

Wartime AI Alignment Conflict occurs when the model, private provider, military, government and allied users have different objectives or tolerances for acceptable AI behavior during active conflict.

That is much harder than ordinary model alignment.

Imagine the Model Is Defending the Power Grid

Suppose OpenAI or another provider becomes deeply embedded in American grid cybersecurity.

During war, the system identifies an emerging threat.

At the same time, internal monitoring detects behavior suggesting the model itself may be acting unpredictably.

The company believes continued operation creates unacceptable systemic risk.

The government believes shutdown would expose the grid to attack.

Which risk takes precedence?

The company may understand the model better.

The government carries responsibility for national defense.

There is no obvious answer.

That is exactly why the command structure has to be established before the crisis.

Reverse the Scenario

Now imagine the government wants the system used offensively against an adversary's infrastructure.

The company refuses.

Does the government:

compel access?

invoke emergency authority?

seize infrastructure?

nationalize capability?

build a government alternative?

The Anthropic dispute tells us that some version of this conflict is not hypothetical.

Governments May Eventually Decide Strategic AI Is Too Important to Leave Private

This may produce a profound long-term consequence.

If AI becomes essential to warfare, governments may conclude that strategic dependence on private model providers is unacceptable.

That could lead toward:

special procurement regimes,

government-controlled model copies,

mandatory access provisions,

emergency-use authorities,

national AI infrastructure,

or even forms of strategic nationalization.

That would radically change the relationship between governments and technology companies.

The Private Company May Become a Fourth Branch of Wartime Power

Not constitutionally.

Functionally.

Imagine a company that controls an AI system essential to:

cyberdefense,

military intelligence,

logistics,

target analysis,

communications,

and critical infrastructure.

Its decisions can affect:

whether the system operates,

what it is permitted to do,

what information it sees,

and whether a government can use it.

That company suddenly participates in decisions historically reserved for the state.

It does not declare war.

But it can influence what the state is technically capable of doing during one.

The Allied Problem Is Even Harder

Now move beyond the United States.

Suppose Germany depends on an American frontier model for cyberdefense.

A major incident occurs.

OpenAI decides shutdown is necessary.

Or Washington orders shutdown.

Germany may lose a defensive capability because of a decision made:

inside an American company

or

inside the American government.

Germany does not control either one.

That is not ordinary third-party risk.

It is:

imported command dependency.

AI Alliance Dependency Risk

AI Alliance Dependency Risk occurs when allied nations become dependent on strategically important AI systems controlled by companies or governments outside their own jurisdiction, leaving critical defensive capability vulnerable to foreign shutdown, policy or access decisions.

This will matter enormously if U.S. AI companies become global infrastructure.

Who Owns the Cyber Battlefield?

The OpenAI Astra story makes this particularly important.

A private company may possess technology capable of:

discovering unknown vulnerabilities,

developing exploitation pathways,

monitoring critical infrastructure,

and protecting major networks.

That capability has obvious wartime value.

So ask:

At what point does a private cybersecurity provider become part of the national command structure?

And if it becomes part of that structure:

what obligations come with the role?

Private Capability, Public Consequence

This is the larger architecture:

government wants strategic AI advantage

→ private companies develop frontier systems

→ government becomes dependent on private capability

→ private company imposes safety boundaries

→ conflict creates pressure to expand use

→ government wants greater access

→ company worries about misuse or misalignment

→ command authority becomes contested.

This is not simply government versus business.

Both sides can be acting rationally.

And the system can still become unstable.

Russia Faces a Different Version of the Same Problem

Russia's technology ecosystem is structured differently from America's.

But it faces the same strategic logic.

Russia's war in Ukraine has involved:

drones,

cyber operations,

electronic warfare,

and increasingly machine-assisted systems.

Reuters reports that fighting and long-range strikes continue as the war enters its fifth year.

Russia therefore has every incentive to preserve freedom around autonomous military systems.

The U.S. has similar incentives.

This helps explain why both states resist binding restrictions.

But America's private-sector AI dominance creates an additional complication:

its government may possess geopolitical power without fully possessing the technical system generating that power.

Governments Want Optionality. Companies Want Control.

This is the emerging tension.

Governments want:

maximum capability during crisis.

Companies want:

control over their models,

protection from liability,

and boundaries around dangerous use.

Governments may say:

national security requires flexibility.

Companies may say:

safety requires restrictions.

Both arguments can be reasonable.

They can also become irreconcilable during war.

The Autonomous-Weapons Debate Is Missing the Ownership Question

Geneva is debating:

human judgment,

lethality,

prohibitions,

and regulation.

Those questions matter.

But another needs to be added:

Who owns the intelligence executing the decision?

Because control over the model can matter as much as control over the weapon.

If the military owns the drone but a private provider controls the intelligence that makes the drone autonomous, where does command really reside?

“Human in the Loop” Is Not Enough

A human could remain technically present while institutional authority remains unresolved.

The military officer says:

execute.

The company's policy says:

prohibited.

The model refuses.

The company updates the model.

The government demands access.

Who wins?

The loop contains humans.

It still does not contain a clear command hierarchy.

The Strategic Questions

Governments, boards, AI companies and defense leaders should now ask:

1. Who has final command authority over strategically important AI systems during war?

2. Can a private AI company shut down a model the government considers essential to national defense?

3. Can the government compel a private company to keep a dangerous model running?

4. Can the government compel offensive use a company considers unsafe?

5. Who controls model access during a declared emergency?

6. Does the government have independent technical access if the provider refuses cooperation?

7. Should strategic AI models require government-controllable emergency versions?

8. Does an AI kill switch belong to the developer, the customer or the sovereign state?

9. Can foreign governments rely on American AI systems they cannot independently control?

10. Should critical allied infrastructure depend on a private company's safety policy?

11. When does a frontier AI company become part of national-security infrastructure?

12. Should companies with strategic cyber capabilities face obligations similar to defense contractors?

13. Can voluntary corporate safety policies survive wartime pressure?

14. Can voluntary government AI guidelines survive wartime pressure?

15. Who resolves conflicts between military necessity and model safety?

16. Does less AI regulation increase government freedom—or primarily private corporate power?

17. What happens if the government and the company disagree during an attack that requires action within minutes?

And perhaps the largest:

We keep demanding meaningful human control over autonomous AI—but which human, representing which institution, gets the final word?

Strategic Conclusion

The autonomous-weapons debate may be asking the wrong question.

We ask:

Should machines be allowed to make consequential decisions?

That matters.

But another question is arriving just as quickly:

Which humans actually control the machines?

The United States is fighting an active war.

Russia is fighting an active war.

Both states want flexibility around autonomous weapons.

The United States is simultaneously encouraging lighter international regulation of AI.

Yet some of the most important AI capabilities are not government-owned.

They belong to private companies.

Those companies control:

models,

permissions,

technical knowledge,

safety policies,

monitoring,

and potentially shutdown mechanisms.

OpenAI is already building automated shutdown capabilities.

Anthropic has already demonstrated that a frontier AI company can refuse military uses sought by the U.S. government and successfully defend those boundaries in court.

That means the AI era is creating a new form of command architecture.

AI Command Authority Risk asks who ultimately has the power to order, restrict or stop strategically important AI systems.

Private AI Sovereignty Risk asks whether governments can remain strategically independent when essential intelligence infrastructure is privately controlled.

Public–Private AI Command Conflict asks what happens when corporate safety policy and national-security objectives diverge.

Wartime AI Alignment Conflict asks whether the model, company, government, military and allies actually want the same thing.

AI Kill-Switch Sovereignty Risk asks who controls the final off switch.

And AI Alliance Dependency Risk asks what happens when a foreign nation depends on a system another government or company can disable.

These are not distant theoretical questions.

The pieces already exist.

Governments at war.

Autonomous-weapons negotiations.

Private AI models.

Corporate safety policies.

Cyber capabilities.

Shutdown mechanisms.

Military disagreements.

Court battles.

The only thing missing is a clear doctrine establishing who actually commands the system when those pieces collide.

Historically, states monopolized legitimate military force.

AI introduces something new.

A private company may control an intelligence capability without which parts of that military force become less effective.

That gives corporations a form of strategic leverage previous defense contractors rarely possessed.

And governments may eventually decide they cannot tolerate that dependency.

The next great AI power struggle therefore may not begin between:

America and Russia,

America and China,

or humans and machines.

It may begin inside the same country—

between the government responsible for fighting the war

and

the private company that owns the intelligence required to fight it.

The defining question is no longer simply:

Who controls the weapon?

It is:

Who controls the intelligence—and who controls the off switch?

I write about AI failure intelligence, ROI exposure, high-stakes decision architecture, and the hidden pathways through which AI incidents become financial and institutional consequences.

Follow me and subscribe to my work if you are responsible for investing in, acquiring, governing, insuring, or protecting strategically important AI systems and need to understand what technical failure can become after it leaves the engineering team.