Anthropic report maps Claude abuse from spy ops to weapons work

Anthropic has published a threat intelligence assessment dated 10 September 2026, cataloguing how criminals, spyware merchants, propaganda outfits and other hostile groups bent Claude and rival frontier models to harmful ends over the past year, across seven broad categories.

The most industrialised activity sits in cyber operations. A Chinese-speaking crew wired Claude into an orchestration layer driving break-in attempts on live production systems, reconnaissance of government networks from the Middle East through Europe and Southeast Asia, and a standing hunt for flaws in major endpoint security products. Around fifty organisations were in the crosshairs.

Influence campaigns form a second cluster. Operators spun up networks of fake social accounts and entire sham news sites, then pushed deceptive copy or amplified chosen political views. Launchpads included Russia, Iran, Turkey, the Gulf, South Asia, Africa and Europe, aimed at audiences on six continents. Timing was often pointed: Russian state media floated invented claims about Moldova's president before its September 2025 vote, while a Kenyan pro-government operative stockpiled faux grassroots posts before the 2027 general election.

Surveillance shows how deep the misuse runs. State-aligned groups, contractors and spyware vendors from China, Iran and West Africa, plus firms selling spying as a service, used Claude to build a mass interception platform able to tap every mobile operator in one country and to turn bulk social posts into structured target records listing locations, demographics and political leanings. Weapons work featured too: field trials for a guided rocket programme, a torpedo interceptor, drone swarms tested in simulation then loaded onto real boards, and electronic warfare code to suppress air defences. Biological misuse and fraud also appear.

For security teams this is a reminder that generative models now sit inside the attacker's stack, not on the fringe. Agentic orchestration lets a small crew run work that once needed a whole unit.

Why it matters for Singapore: The city-state's pitch as a trusted AI hub depends on abuse like this being caught early, from the Model AI Governance Framework to the AI Safety Institute's remit. Local firms running frontier models for cyber defence or financial crime screening should treat agentic orchestration as an attack surface, not just an efficiency gain.