Anthropic says people overseas used Claude to help track U.S. naval forces, develop weapons, control groups of drones and figure out why a guided rocket failed after it was test-fired.
For years, the argument over artificial intelligence and warfare has centered on what might happen someday.
Someday has arrived.
Anthropic, the company that makes the Claude artificial intelligence system, says it has discovered people connected to foreign governments and militaries already using Claude to help find and track American naval forces, develop weapons, control groups of drones and build systems designed to interfere with an enemy’s radar and communications.
In one case, a group in Yemen used Claude while developing the software needed to guide a rocket.
They built it.
They took it into the field.
They fired it.
It failed.
Then they went back to Claude to help figure out what went wrong.
That may be the simplest way to understand what has changed.
This wasn’t a Pentagon exercise about what artificial intelligence might be capable of five or 10 years from now.
According to Anthropic, this was happening now.
In a new report covering activity it discovered between December 2025 and August 2026, Anthropic describes six cases in which Claude was used for work involving conventional weapons or military systems. Three were connected to China, two to Russia and one to Yemen.
The company also discovered something particularly important to Americans serving in the Middle East.
People connected to Iran were using Claude to help develop information that could be used to target U.S. naval forces.
Finding American warships
Anthropic says an Iran-linked group used Claude to develop recommendations for targeting U.S. naval forces in the region.
They weren’t simply asking Claude where an American ship was located.
They were building a system to gather bits and pieces of information available on the internet and turn them into something that could be used against American forces.
The group gathered names of U.S. military personnel from captions on military photographs.
It collected identifying signals transmitted by ships and aircraft.
It developed ways to search commercial satellite photographs.
And it assembled a list of public websites that could help reveal where American naval forces were operating.
The group also used Claude to research possible weaknesses in equipment aboard U.S. ships, including satellite communications, computer networks and the computerized equipment used to control machinery.
None of those pieces of information necessarily means much by itself.
Put them together, however, and they can begin to provide a picture of where American forces are, who is with them and where their equipment might be vulnerable.
Finding and assembling that information once required people to search through large amounts of material.
Artificial intelligence can help do it much faster.
Anthropic shut down the account and said it gave information about what it found to government authorities.
Building a guided rocket
The case in Yemen went beyond gathering information.
According to Anthropic, a group used Claude while developing the computer software that tells a guided rocket where it is, where it needs to go and how to steer itself there.
The group adapted software that was already publicly available and worked to make it function with relatively inexpensive flight-control computers.
Then the project left the computer screen.
The group test-fired a guided rocket.
It didn’t work.
They returned to Claude with information from the test and used the AI to help determine what had gone wrong.
Anthropic does not say Claude successfully helped them produce a working weapon.
But whether that particular rocket worked isn’t really the most important part of the story.
Look at the process.
Design the weapon.
Write the software.
Build it.
Test it.
Study the failure.
Try to fix it.
Artificial intelligence had become part of that process.
Drones that work together
Anthropic says people connected to Russia used Claude for another kind of weapon: groups of drones designed to work together.
Instead of one person controlling one drone, the goal was to have multiple drones coordinate their movements and attacks.
Claude was used to help develop the software.
According to Anthropic, the project progressed beyond computer simulations and was eventually loaded onto actual equipment.
The system was being designed so drones could choose targets and trigger their explosives without a person making each individual decision.
Combat video from the war in Ukraine was also used to help train the system.
That distinction is important.
For years, much of the debate about artificial intelligence and warfare has centered on the prospect of machines someday deciding for themselves whom to kill.
Anthropic’s report points to a more immediate problem.
Humans don’t have to hand control of a war over to artificial intelligence for AI to change warfare.
They can use it as an engineer, programmer and research assistant while they build the weapons themselves.
China and Taiwan
Anthropic also found people connected to China using Claude for military work.
One case involved software designed to help disrupt an enemy’s radar, communications and air defenses.
Claude helped with repeated versions of a system that examined military targets such as radar stations, anti-aircraft missile batteries, communications sites and military command centers.
The software helped determine which targets were most important, which were most vulnerable and which should be attacked first.
Then the targets became specific.
The person using Claude changed the exercise to 12 locations in Taiwan.
They included air bases, an early-warning radar installation, a military command bunker, American-made Patriot air-defense systems and Taiwan’s own Tien Kung anti-aircraft missile systems.
Anthropic said the person appeared to have connections to Chinese military research, including the People’s Liberation Army’s Academy of Military Sciences.
In another China-linked case, Claude was used for work involving a system intended to protect ships against torpedoes.
A third involved gathering information about advanced weapons, including weapons that use concentrated energy rather than conventional explosives, and about U.S. military technology.
Anthropic does not claim every project became a working weapon.
But taken together, the cases show something that is difficult to dismiss.
Artificial intelligence is no longer simply something militaries are considering for the future.
People are using it for military work now.
What AI changes
Claude didn’t invent missiles.
It didn’t invent drones.
It didn’t invent spying on enemy ships.
Engineers, programmers and intelligence services have been doing those things for generations.
What artificial intelligence can change is how much expertise someone can call upon and how quickly they can use it.
A programmer who gets stuck on part of a weapons system can ask AI for help.
Someone trying to locate American ships can give AI thousands of pieces of publicly available information and have it help sort through them.
An engineer can bring back the results of a failed weapons test and ask the AI what went wrong.
And a country or armed group without a large staff of engineers and specialists may be able to do work that once required more people, more expertise and more time.
That may be one of the most important military consequences of commercially available artificial intelligence.
The United States can spend billions of dollars developing specialized military AI.
Someone on the other side may be able to begin with a laptop.
That doesn’t mean a novice suddenly becomes a missile engineer.
The failed rocket in Yemen is evidence of that.
But weapons development has always involved failure.
Engineers build something.
They test it.
It breaks.
They determine why.
Then they build another one.
Now AI can help with that process.
The danger isn’t limited to conventional weapons
Anthropic also found five cases involving biological research that it concluded could potentially contribute to the development of biological weapons.
One involved research on chikungunya, a virus spread by mosquitoes.
The researchers were studying ways of changing the virus that could affect how easily it spreads and how well it escapes the human immune system.
The proposed work included genetically altering the virus and testing different versions in animals.
Anthropic became especially concerned because the work was intended to take place at a military research institute.
There is an important caution here.
Anthropic does not say it proved that the researchers were trying to build a biological weapon.
The same type of research can have legitimate medical uses, including understanding diseases and developing vaccines and treatments.
That is also what makes it dangerous.
Research that can help scientists fight a disease can sometimes provide knowledge that could be used to make one more dangerous.
Anthropic shut down the accounts and reported what it found to authorities.
America uses AI, too
None of this is happening in a vacuum.
The United States military is developing and using artificial intelligence as well.
AI can help military commanders and analysts sort through enormous amounts of information. It can help operate unmanned aircraft and other machines, move supplies and equipment, and identify potential targets.
That creates a problem the United States has encountered before.
Military technology doesn’t remain on one side for long.
Radar didn’t.
Jet aircraft didn’t.
Guided weapons didn’t.
Drones didn’t.
Artificial intelligence won’t either.
But AI may spread much faster.
A country that wants to build a modern fighter aircraft needs factories, specialized materials, trained engineers, thousands of workers and billions of dollars.
Access to an artificial intelligence system is considerably easier to obtain.
Anthropic says some of the people it discovered tried to get around the company’s safety restrictions by breaking complicated projects into smaller requests that appeared less dangerous.
Anthropic closed the accounts and says it has strengthened its protections.
But Anthropic isn’t the only company building artificial intelligence.
And the technology isn’t going away.
That leaves the United States and every other military with a new reality.
The same artificial intelligence that can help an American programmer write software or help a college student understand physics can also help someone search for an American warship.
It can help someone write software for a weapon.
It can help someone test that weapon.
And when the weapon fails, it can help them figure out why.
For years we’ve been asking what warfare will look like when artificial intelligence reaches the battlefield.
We don’t have to wonder anymore.
AI isn’t coming to the battlefield.
It’s already there.