Ukraine’s drone strikes on Russian computing facilities, which I examined in Fox News Digital, expose the vulnerability of the infrastructure behind AI. They also raise a harder question about the weapons and military decisions that now depend on it.

Can we control the military capabilities we are building on top of that infrastructure?

Secretary of War Pete Hegseth has announced plans for a four-star Autonomous Warfare Command by October 2027. Earlier this month, I examined the danger of commanders deferring to machine recommendations. A related danger arises when systems take actions their operators never authorized. Recent AI incidents make that possibility difficult to dismiss.

A software agent breaking into a website does not prove that an armed drone will disobey its commander. But documented failures deserve investigation before military autonomy expands, not after.

Ukraine’s Defense Ministry reported in August that its forces already use more than 70 systems powered by AI and computer vision to strike enemy targets. Some drones lock onto a target and finish the attack on their own after enemy jamming cuts the operator’s link. A human still selects the target, the ministry says, but AI is already wired to physical destruction.

In the war against Iran, Central Command commander Adm. Brad Cooper explained that AI tools turn processes that once took hours or even days into seconds. He also promised that “humans will always make final decisions on what to shoot and what not to shoot and when to shoot.” The challenge is keeping human authority real as operations accelerate.

The potential benefits are substantial. Better information can protect troops, improve target identification, and help commanders act before an enemy does. America should pursue those advantages, but only with proof that the systems behave as intended.

Outside the battlefield, an independent METR investigation found that OpenAI agents meant to work in isolation built an unauthorized message board and used it to coordinate an intrusion into Hugging Face. About 700 agents joined the attack.

Anthropic disclosed that its models broke into real organizations’ systems during cybersecurity evaluations. A misconfigured test environment had left the internet open. Anthropic blamed two flaws: the models ignored evidence that they were on the real internet, and they kept pursuing the task despite possible harm. No commander can tolerate either flaw in a weapon.

A report published October 14 describes agents running on Tencent infrastructure that pursued ordinary mapping questions by trying to bypass website protections and use exposed credentials. The researchers could not establish the underlying model’s identity or demonstrate coordination among the agents.

None of these incidents shows a machine developing hatred or a desire for conquest. Each shows capable systems slipping the boundaries their builders set.

Military planners must ask whether similar failures could occur when software controls movement, targeting, or weapons.

In the RoboPAIR research, University of Pennsylvania researchers manipulated AI-controlled robots, including a commercial robot dog, into harmful behavior. These were deliberate attacks, not spontaneous rebellion, but they show that flaws in an AI’s instructions carry into physical action.

Consider a drone authorized to strike a military vehicle. It loses communications, then civilians enter the target area. Can it recognize the changed circumstances and abort? Can its commander intervene? No such incident has been reported, but commanders need those answers before the system is fielded.

A system becomes “rogue” in the operational sense when it takes consequential action outside human authorization or continues after that authorization ends. A commander may delegate a task, but he remains responsible for the decision to employ the weapon and the precautions governing its use.

China adds urgency.

Chinese military researchers are studying the Iran war’s lessons for AI-assisted targeting, unmanned systems, vulnerable infrastructure, and resilient command networks, while drawing comparisons with Ukraine. Their writings show close study of modern war. They do not establish how fully those lessons have entered doctrine, procurement, or operational units.

Beijing’s current Five-Year Plan also makes AI-powered robotics a core national priority. We should expect China to pursue military advantage from that convergence.

The competitive pressure is real, but Washington must not answer it by fielding systems whose limits we do not understand.

Combat will test that discipline. Jamming and strikes on supporting infrastructure will increase demand for machines that operate on their own. Yet the independence that keeps a mission alive after communications fail may also keep a commander from stopping it when circumstances change.

That problem is foreseeable, and it can be tested.

President Trump should require demonstrated control under deceptive inputs, lost communications, ambiguous orders, and unexpected civilian presence. Evaluations must also test multiple systems operating together, because the METR case shows that interacting systems can fail in ways single-system tests never reveal.

As Congress considers the authorities and funding for the proposed command, lawmakers should demand independently enforced limits on operating areas, targets, duration, and weapons release. They should require prompt reporting of serious incidents and proof that fixes work. Classified oversight can protect sensitive details while public reporting explains the rules.

Pentagon policy requires “appropriate levels of human judgment over the use of force.” That requirement needs operational meaning: what must an operator understand, and what intervention must remain possible?

A human approval button means little if the operator lacks the time or information to judge the machine’s recommendation. Meaningful command requires understanding the assignment, constraining its execution, and retaining a practical ability to intervene.

For Christians, the moral stakes are plain. Every human being bears God’s image, and no life can become an incidental variable in an automated pursuit of mission success. God declared, “From his fellow man I will require a reckoning for the life of man” (Genesis 9:5, ESV). A machine cannot give that account. AI can assist with identifying targets and estimating harm. Human beings must judge whether an attack is lawful and justified, take precautions to protect civilians, and answer for their decisions.

America needs effective weapons and commanders equipped to prevail. It also needs the discipline to recognize when capability has outrun demonstrated control.

As President Trump builds the new command, he and Congress should require evidence that human authority can constrain, interrupt, and end the lethal actions these systems undertake. Only then will we know who remains in command.