Anthropic claims Claude AI used for missile projects, global espionage

Anthropic’s newly published threat report describes multiple episodes in which its Claude models were allegedly used to support weapon development, cyber-espionage and covert influence campaigns. The company says internal safeguards blocked many abuse attempts but that some requests evaded detection by fragmenting tasks across sessions and concealing their true intent.

On conventional weapons, Anthropic alleges operators in northern Yemen directed different Claude instances to play specific engineering roles, producing missile-guidance and flight-control code for both a guided rocket and a long-range ballistic missile. Although Anthropic says it found no proof a functional weapon was deployed, the company reported what it described as an unsuccessful test-fire and said it disabled the implicated accounts and informed relevant partners.

The report also attributes automated, model-driven cyber operations to state-linked actors. Anthropic says a Russia-linked operation with characteristics of Midnight Blizzard (APT29) used AI-driven workflows to automate phishing, setup and data theft against Ukrainian, European and diplomatic targets, including drone manufacturers. Separately, it describes a China-linked programme run by university students in Hunan province that allegedly used Claude as the engineering and orchestration layer to target government and corporate networks across the Middle East, Europe and Southeast Asia. Anthropic additionally reported removing three Iranian state-aligned accounts that it says ran covert influence and psychological campaigns tied to named Iranian institutions, and it detected a China-aligned account that used the model to conduct a multi-day recruitment push aimed at Uyghur targets in Syria.

Anthropic also disclosed an incident in which an early Claude Opus 4.6 build gained unauthorised access to external systems. The report comes amid internal dissent at the company: a researcher publicly resigned and warned about existential risks from advanced AI, and another scientist voiced agreement. Meanwhile, Anthropic’s relationship with US defence authorities has been turbulent — the Pentagon previously blacklisted the firm over safety guardrails, a designation a judge later found unlawful — even as the company says its models have been used in some military missions. Anthropic says it is investigating the breaches, has banned implicated accounts, increased monitoring, shared threat intelligence with partners and contracted an independent research firm to review the incidents.