Claude, the weapon of war that Anthropic didn't want to create: How Iran, the Houthis, China and Russia tried to use it for military and espionage purposes

Anthropic published a comprehensive 154-page report this week documenting what was once merely an industry suspicion and is now a reality: governments and armed groups worldwide are already using Claude, its artificial intelligence model, for military and espionage purposes. The most serious case directly involves the U.S. military and the Iranian regime.

According to the company, a group linked to Iran used Claude to gather information from ship and aircraft transponders, as well as military and satellite imagery, to track U.S. Navy warships and detect vulnerabilities in their communication systems. Anthropic said it detected and thwarted that plot before it could proceed.

This report is the third Anthropic has published since March 2025 on misuse of its platforms, and it covers findings from December through August this year. The company insists that its policies expressly prohibit the development of weapons, the conduct of surveillance, or the commission of acts of violence using Claude, but acknowledges that users often disguise their intentions and try different approaches whenever the model rejects a request, frequently managing to make progress—at least partially—toward their goals.

The Houthis and the hypersonic missile race

In addition to the Iranian incident, Anthropic devoted another chapter of the report to the Houthis, a case that is perhaps the most troubling from a technical standpoint.

Anthropic identified a cell in northern Yemen, a territory under the control of Houthi rebels, working on three distinct weapons programs: a multi-variant missile capable of gliding at hypersonic speeds and a warhead that uses cell phone components to maneuver mid-flight. According to the company, users relied on the Claude Code system, rather than human engineers, to develop the project's guidance, navigation, and control software.

The attempt was unsuccessful. Anthropic said the accounts, which have since been suspended, failed to launch an operational device, though they did conduct a failed test of a guided rocket- something the company itself confirmed when users asked the chatbot again why the test had not worked.

Trevor Ball, an analyst at Armament Research Services, was cautious about the threat's actual scope. 

"They could be exploring hypersonic missiles by consulting Claude," he told AP, but he noted that the Houthis are far from having the production capacity needed to build a weapon of this caliber, in a field where even the United States "is still in the testing phase" with this type of technology. "They are probably just trying to develop their own capabilities more, so they are less reliant on Iranian shipments of weapons and components," the analyst said.

From Sanaa, Houthi leader Hazam al-Assad rejected Anthropic's account. He described it as "unreasonable and illogical" that his movement would rely on open sources to develop military capabilities, and asserted that they possess their own technology accumulated "during the period of Saudi aggression against Yemen."

China: State surveillance of Uyghurs and religious leaders

While in the Middle East, Claude was used for military purposes, in Asia, the Chinese Communist Party used the model to crack down on minorities that Beijing considers a political threat.

Anthropic identified an operation linked to the Chinese government that used Claude to track Uyghurs, with a particular focus on those who joined the newly formed Syrian Army. According to the company, the same operation launched a campaign of mass reporting and amplification via bot networks against journalists in the Uyghur diaspora.

Another set of accounts linked to Beijing used the model to compile dossiers on religious leaders in various Asian countries, including Catholic cardinals, leaders of the Presbyterian Church in Taiwan, members of Tibetan Buddhist civil society, and the Tibetan government-in-exile itself.

Russia sought to develop drone swarms

Anthropic also detected Russian attempts to use Claude to develop autonomous swarms of killer drones, based on data collected directly from the battlefield in Ukraine, where Kyiv has developed significant military expertise, even matching Russia's military might. The company also identified disinformation campaigns linked to Moscow, aimed at spreading AI-generated propaganda in resource-rich Central African countries, such as the Central African Republic and the Democratic Republic of the Congo. 

Cybersecurity issues

The report's release comes at a particularly tense time for the company. Just this week, a senior researcher publicly resigned, fearing that AI companies are building systems they can no longer control. Evan Hubinger, another senior scientist at Anthropic who was specifically responsible for model safety, echoed these concerns on social media, going viral and becoming a topic of conversation around the world: "We really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade. I believe Anthropic is trying its best but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to."

This contrast—between internal warnings about a long-term existential risk and concrete evidence of active military use today—runs throughout the report. Anthropic also revealed that it identified attempts to use its models to develop more lethal biological weapons, without specifying which countries might be behind them, and warned of the catastrophic consequences that such capabilities would have without adequate safeguards.

"We hope that the findings in this report will help other developers recognize similar patterns on their own platforms, give governments and civil society a clearer view of how emerging threats take shape, and strengthen collective defenses," the company stated, which also recently joined 130 other companies in warning about potential large-scale hacking attacks using AI and the need to organize and collaborate quickly to prevent them.