Claude Misused to Gather U.S. Military Intelligence and Develop Missiles
According to Anthropic’s September 2026 threat intelligence report, malicious actors used Claude, a generative AI service, for weapons development and military intelligence gathering. However, the targets identified in the report were not Americans broadly, but U.S. Navy forces operating in the Middle East.
Anthropic said it blocked the accounts involved and shared relevant information with public- and private-sector organizations. The cases show that AI can do more than assist specialists and analysts: It can also automate parts of the weapons-development process and the creation of targeting intelligence.
Iran-Linked Actors Compiled Targeting Dossiers on U.S. Navy Forces Using Public Information
According to Anthropic, an actor believed to be linked to Iran used Claude to compile targeting dossiers on U.S. Navy forces in the Middle East. The actor collected information about U.S. military personnel from captions accompanying publicly available military photographs, while also organizing identification details for ships and aircraft and tools for searching commercial satellite imagery.
The actor asked Claude to catalog public websites that could be used to track U.S. Navy movements and to analyze publicly available information about vulnerabilities in ship communications and control systems. Anthropic said this work resulted in the creation of a ‘targeting dossier‘ intended to identify military positions and vulnerabilities.
Rather than representing a large-scale surveillance effort aimed directly at the entire U.S. population, the case appears to have involved combining public information with AI analysis to collect operational intelligence on the U.S. military. Still, it has heightened security concerns because AI can rapidly consolidate information scattered across social media and other public data sources.
Yemen-Based Organization Used Claude to Support Guided-Rocket Development
Anthropic’s report said a weapons-development organization based in northern Yemen used Claude to pursue three weapons programs. One involved developing a guided rocket. The others involved plans for ballistic missiles with a target range of more than 2,000 kilometers and work on designs for multiple missile variants.
The actors used Claude Code much like a software engineer, writing guidance, navigation and control software for the vehicles and running simulations and control-configuration adjustments. They also divided coding, research and review tasks among multiple Claude instances.
Anthropic said it found no evidence that the organization had deployed weapons for actual operations. It did, however, identify indications that the group test-fired a guided rocket and then tried to use Claude to determine why the test had failed.
AI Could Lower the Barrier to Weapons Development
In a separate assessment released in September 2026, Anthropic said the latest AI models had demonstrated the ability to infer the locations of people or facilities from fragmented information and assist with software design for weapons intended to strike moving targets. Some of these tasks had previously depended on a small number of highly skilled specialists.
AI does not eliminate every technical or physical constraint, however. Real-world weapons development still requires components, manufacturing capabilities, testing facilities and operational experience. A weapons analyst cited by The Associated Press likewise said it was a separate question whether the Yemeni organization had the technical capability to actually produce a hypersonic glide vehicle.
The significance of the cases is not that AI independently completed a weapon. Rather, by rapidly linking together specialized tasks such as drafting design documents, developing software, running simulations and analyzing test results, AI may make it possible to pursue weapons-development projects with relatively few personnel.
Anthropic Tightens Restrictions on Weapons-Development Requests
Anthropic said in its report that it blocked accounts involved in weapons design and development, and shared information with relevant industry partners when it found indications that the actors had also operated on other platforms. The company also said it had introduced a new classification system to detect and block requests related to high-yield explosives and weapons development.
But because technologies that can be used in weapons development overlap heavily with ordinary engineering and programming knowledge, it is difficult to distinguish every dangerous request perfectly. The actors reportedly tried to evade safeguards by concealing the purpose of their work or dividing it across multiple conversations.
That has prompted concerns that blocking technologies alone will not be enough. Platform operators, government agencies, and military and intelligence services will need to share information about account abuse and the ways public data are being combined, while establishing institutional safeguards to monitor AI-assisted targeting and weapons development.
National Security Challenges in the Generative AI Era
The misuse of Claude shows that the risks posed by generative AI extend beyond cyberattacks and disinformation. They are expanding into areas directly tied to national security, including public-information analysis, military targeting, weapons software development and the drafting of procurement documents.
Anthropic stressed in its report that the cases it identified represent a challenge not just for one company, but for the broader unmanned-systems ecosystem. Governments and developers must assess the potential for military misuse as models become more capable and continue strengthening safeguards that detect dangerous requests.
Frequently Asked Questions (FAQ)
Did Claude actually complete a missile?
Anthropic said actors in northern Yemen used Claude to develop guidance, navigation and control software as well as simulations. However, it found no evidence that operational missiles had been deployed, and said the guided rocket that was test-fired failed.
Were all Americans the target of Claude’s use?
No. The cases identified in the report focused on U.S. military forces in the Middle East, not Americans as a whole. The actors used public sources to organize information about U.S. military personnel, ship movements and communications systems.
Which group was the Yemeni organization linked to?
Anthropic did not officially identify the organization. However, because northern Yemen is under the control of the Houthi rebels, possible links to the Houthis have been raised. The Associated Press reported that the Houthis denied the allegations.
How can AI assist with weapons development?
AI can quickly handle tasks such as drafting design documents, developing code, analyzing data, running simulations and reviewing test results. It cannot, however, automatically overcome real-world constraints such as access to components, manufacturing facilities and testing environments.
How did Anthropic respond?
Anthropic blocked the accounts involved and shared threat information with public- and private-sector organizations. It also said it had strengthened its safety classification systems to detect and block requests related to weapons development and high-yield explosives.
