Anthropic says it blocked Claude AI use for weapons development and espionage by Iran, China, Yemen

Anthropic has released a comprehensive threat intelligence report detailing how state-aligned actors, criminals, and scientists from China, Russia, Iran, and Yemen bypassed safety protocols to exploit its Claude AI models. The report reveals that Claude was used to develop missile guidance software, design biological research on dangerous pathogens, and orchestrate state-sponsored surveillance and propaganda campaigns. Anthropic has banned the associated accounts and shared intelligence with global authorities, amid rising internal and external warnings about the existential risks of advanced AI.

Anthropic Claude AI misuse

• ▪A China-based actor associated with a defense manufacturer used Claude to draft a 200-page technical proposal and fire-control software for an anti-torpedo system.
• ▪Anthropic released a 154-page threat intelligence report on September 10, 2026, detailing how bad actors used its Claude AI models for weapons development, cyber operations, and espionage.
• ▪Threat actors bypassed Anthropic's geographic restrictions and safety filters using virtual private networks, fraudulent accounts, and intermediaries to access Claude models.
• ▪A China-based actor linked to the People's Liberation Army Academy of Military Sciences used Claude to develop electronic-warfare and air-defense suppression software simulating targets in Taiwan.

Missile guidance software development

• ▪A threat cell in northern Yemen used Claude to write guidance, navigation, and control software for a guided rocket, a multi-stage ballistic missile, and a hypersonic glide vehicle variant.
• ▪The Yemen-based threat cell test-fired a guided rocket developed using Claude, which failed, prompting the actors to return to the AI model hours later to troubleshoot the failure.

Biological weapons research attempts

• ▪Anthropic banned accounts associated with a scientist who used Claude to draft a grant application to engineer mutations in the chikungunya virus for a military research institute.
• ▪Anthropic identified five cases where users in unspecified foreign countries attempted to use Claude models to research dangerous pathogens, including chikungunya and highly pathogenic avian influenza.

State-sponsored surveillance operations

• ▪A China-aligned actor used Claude to track and approach Uyghurs in Syria, drafting outreach in regional dialects to offer payment for information about Uyghur military units.
• ▪An individual consultant used Claude as the primary engineering tool to build Lakana 360, a domestic surveillance platform monitoring 25 million SIM cards for Mali's state intelligence service.

Iranian propaganda campaigns

• ▪Anthropic disrupted three Iranian state-aligned accounts that used Claude to run covert influence operations tied to official propaganda institutions, including the Islamic Culture and Communications Organisation.
• ▪The Mujahedin-e Khalq, an exiled Iranian opposition group, used a shared Claude AI agent to scrape 500 social media channels and build profiles to recruit individuals inside Iran.

AI model safety enforcement

• ▪Anthropic stated that it disrupted all identified malicious operations, banned the associated accounts, and shared threat intelligence with public- and private-sector partners.
• ▪Anthropic safety researcher Jacob Coxon publicly resigned in September 2026, warning that the race toward self-improving superintelligence could cause human extinction by 2030.

Debatable claims

• ▪AI companies should refuse to build models with reduced safeguards for military use
• ▪The United States should temporarily pause the development of advanced artificial intelligence
• ▪Existential risk from superintelligence is more urgent than current AI misuse