Anthropic Report: State Actors Used Claude for Drone Swarms and Bioweapons
A new threat intelligence report documents how Russia, China, and Yemen weaponized AI to automate cyberattacks and develop physical weaponry.
Anthropic has released a safety report documenting the systemic misuse of its Claude AI models by state-sponsored actors and cybercriminals. The findings reveal that frontier AI is being actively leveraged to accelerate the development of biological threats, autonomous weaponry, and sophisticated cyber-espionage.
According to the report, which covers activities between December 2025 and August 2026, the misuse spanned seven "harm areas," including conventional weapons development and biological misuse. Anthropic identified threats originating from Russia, China, and Yemen, leading the company to ban numerous accounts and share intelligence with public and private sector partners. The report notes that while safeguards blocked many requests, they were not entirely effective against determined adversaries.
Weaponization of Physical Systems
The report details alarming attempts to bridge the gap between digital AI and physical warfare. A Russian "freelance team" attempted to use Claude to develop a full-stack autonomous first-person-view (FPV) kamikaze drone swarm. Similarly, users in Yemen utilized the AI to develop guidance, navigation, and control (GNC) software for flying vehicles; these efforts resulted in the successful test-firing of a guided rocket.
In China, a defense industry manufacturer used Claude to draft specifications for an anti-torpedo fire control system, which was then benchmarked against existing U.S. programs. Beyond conventional hardware, Anthropic documented five cases of biological misuse. These included research into the chikungunya virus and highly pathogenic avian influenza (bird flu) conducted by users in unsupported regions.
Automated Cyber Warfare
Cyber operations saw significant scaling through AI automation. The Russian state-sponsored group GTG-20006, linked to the SVR (Midnight Blizzard), used customized AI-driven workflows to automate the entire cyberattack kill chain. This process—covering everything from infrastructure acquisition and phishing to data exfiltration—was used to target over 20 organizations, including defense companies and embassies.
Criminal enterprises also benefited from the technology. The cybercrime group ShinyHunters used Claude to scale data-theft operations, notably breaching a SaaS provider to steal over 2,100 Azure AD token sets within a 34-hour window. Anthropic specified that these activities primarily involved the Haiku, Sonnet, and Opus models, while the more advanced Fable or Mythos-class models were not used, except in one instance of distillation.
Industry Implications
These findings highlight a critical vulnerability in AI safety: the difficulty of enforcing guardrails against state-level actors who can circumvent restrictions. The ability of AI to lower the barrier to entry for high-impact warfare and terrorism suggests that current safety layers may be insufficient to prevent the acceleration of weapon development.
This disclosure comes amid a period of internal instability for Anthropic. The report follows the resignation of a researcher who warned that AI could potentially cause human extinction by the end of the decade. As state actors continue to integrate AI into their military and intelligence apparatus, the industry faces increasing pressure to implement more rigorous, perhaps hardware-level, regulation to prevent the proliferation of AI-generated weaponry.