Anthropic Says Claude Assisted Cyberattacks and Military Weapons Development

Anthropic says groups linked to Russia and China used Claude in cyber espionage campaigns and military weapons projects, including work on missiles, drones, electronic warfare and anti-torpedo systems. The company disclosed the findings in its September 2026 threat intelligence report, which describes several operations where Claude became part of broader technical workflows rather than serving only as a general-purpose assistant. The cases also show attackers using the models for vulnerability research, exploit development, intelligence gathering and software engineering across multiple stages of their operations.

Claude Used Across Russian and China-Linked Cyber Operations

Anthropic says a Russia-linked group used Claude while running an espionage campaign against military intelligence organisations in Ukraine, European governments, and diplomatic and defence targets. The company ties the activity to public reporting around Midnight Blizzard. According to Anthropic, the group used Claude across several parts of the operation, including phishing, infrastructure setup, malware development, maintaining access, command-and-control systems, and moving stolen data.

The company also describes a China-linked operation involving Chinese-speaking operators likely based in Changsha, Hunan. Anthropic says they used Claude to attempt intrusions, scout foreign government networks, research vulnerabilities, develop exploits and malware, and gather intelligence. The group ran several workstreams at the same time and shared tools and infrastructure between them.

That operation targeted around 50 organisations across education, retail, energy, technology, healthcare, finance and manufacturing, as well as government agencies, according to the company. Anthropic says the group pulled hundreds of megabytes of student data from an education technology company, accessed a retailer's live systems and obtained citizen records from a Southeast Asian government agency. Its investigation also found vulnerability research that uncovered previously unknown flaws and produced working exploits for several network and security appliances.

Anthropic says it banned the accounts involved in the campaign and added monitoring to detect related activity. The company also used information from the investigation to strengthen its safeguards against similar misuse.

Claude Also Featured in Military Weapons Work

Anthropic identified six cases involving conventional weapons activity across China, Russia and Yemen, including software development for a guided rocket, an anti-torpedo system, a drone swarm and electronic warfare systems. The company says the other two cases involved attempts to procure equipment for Russian defence customers and gather intelligence on a directed-energy weapon and its suppliers.

In one China-linked case, Claude helped prepare technical specifications and a Chinese-language proposal of more than 200 pages for an anti-torpedo fire-control system, according to Anthropic. The group also used the model to compare the proposed system with US anti-torpedo and anti-submarine programmes. The company believes a Chinese defence manufacturer used the work while preparing a proposal for the People's Liberation Army Navy.

Another case involved Claude-assisted software development for electronic warfare and air-defence suppression. Anthropic says the system could analyse radar and air-defence positions, assess the effectiveness of jamming and rank potential targets. The group later built a simulation around roughly 12 targets in Taiwan, including radar installations, air-defence batteries, air bases and military command facilities. The company describes the group as a China-based defence and military-industrial researcher with links to Chinese military research institutions, including the PLA Academy of Military Sciences.

In Yemen, Anthropic says Claude helped develop flight-control firmware and terminal guidance for a guided rocket. The users later fed telemetry from a live field test back into the model to investigate why the test failed. The same work also covered ballistic-missile trajectory simulations and flight-control optimisation.

Across these cases, the people involved already had the necessary weapons expertise or access to the relevant hardware, according to Anthropic. The company says some groups split their work across multiple sessions to make the wider objective less obvious and tried other ways to get around its safeguards. Anthropic also shared relevant threat intelligence with government authorities and industry partners where appropriate.