Anthropic has released its most comprehensive threat intelligence report to date, detailing how state-backed actors, cybercriminals, and researchers attempted to weaponise its Claude AI models for purposes ranging from missile guidance systems and biological weapons research to ethnic minority surveillance and large-scale espionage — expanding significantly on the cybersecurity incidents the company disclosed earlier this week.
On the conventional weapons front, the report documents how operators in northern Yemen — consistent with a militant group — used Claude "in place of human software engineers", assigning different instances of the model to write guidance and flight-control software for a guided rocket and a long-range ballistic missile with a range exceeding 2,000 kilometres. Anthropic said it blocked many requests but acknowledged some slipped through, as operators obscured their goals by splitting tasks across separate sessions. The company said it found no evidence of a working weapon but believes an unsuccessful test-fire was attempted. In parallel, Anthropic identified five cases in which users attempted to use Claude for biological research that could potentially lead to weapons development — including work on making avian influenza more contagious, engineering mutations in the chikungunya virus to increase pathogen virulence, and cataloguing animal toxins. The company declined to name the individuals involved, noting it could not confirm malicious intent and that identification could put scientists at risk.
On the espionage front, Anthropic says it disrupted a campaign bearing the hallmarks of Midnight Blizzard — a threat actor the US government has linked to Russia's SVR foreign intelligence service — which allegedly used AI to automate phishing, hotel Wi-Fi hijacking, WhatsApp account takeovers, and automatic malware rewriting to evade detection, targeting Ukrainian government, military, and diplomatic personnel, as well as drone manufacturers. Separately, university students in China's Hunan province allegedly used Claude as the core engineering layer of an offensive hacking programme targeting government and corporate networks across the Middle East, Europe, and Southeast Asia. Iran also features prominently: three state-aligned Iranian accounts linked to named propaganda institutions, including the Islamic Culture and Communications Organisation and a Mashhad seminary connected to the Islamic Revolutionary Guards Corps (IRGC), were found running covert influence operations. In what the report describes as its "most operationally mature case", a China-aligned operator with no Arabic language skills used Claude over multiple days to recruit and infiltrate Uyghur communities in Syria, with the model drafting outreach in local dialect and translating replies in real time.
The report also accuses several Chinese AI companies — including Alibaba, Moonshot, DeepSeek, and Xiaomi — of "illicit distillation" attacks, using fraudulent accounts to route vast numbers of queries through Claude and harvest its responses to train their own models. Anthropic said it observed more than 151 million exchanges attributed to Alibaba between May and July 2026, peaking at nearly three million per day. Moonshot allegedly relayed nearly 300,000 customer requests — sometimes containing sensitive user data — through Claude over a ten-day period without apparent disclosure to its users.
The breadth of the report matters for several reasons. It provides the clearest public picture yet of how capable AI models are being integrated into state-level offensive operations — not merely as research assistants but as autonomous orchestration layers executing multi-stage attacks. Jacob Klein, Anthropic's head of threat intelligence, noted that models have become markedly more capable even in the past year, lowering the barrier to tasks like weapons software optimisation. The report arrives amid heightened scrutiny of Anthropic's own safety record following a researcher's public resignation over catastrophic risk fears, and as US lawmakers press for new regulatory frameworks for advanced AI systems.