Anthropic published its most detailed threat intelligence report to date on Thursday, saying it had disrupted a wide range of attempts to misuse its Claude models. The 154-page document describes cases involving weapons development, biological research, cyberattacks, surveillance and influence operations, and it says the company banned the accounts involved.
The most striking case involved a cell of threat actors based in northern Yemen, a stronghold for Houthi rebels. Anthropic said the group ran three weapons programs: a guided rocket, a multi-stage ballistic missile with a stated range goal above 2,000 kilometres, and a missile with a hypersonic glide vehicle variant. The report says the actors used Claude's coding tools to write the guidance, navigation and control software that steers and stabilises a flying vehicle, work normally done by computer engineers.
Anthropic said its safeguards blocked many requests but not all. The actors hid their goals and the products the software was meant for, and they split their work across multiple sessions so no single conversation revealed their full intent. They also ran several Claude instances at once, giving each a role, in what the report compares to how a lead delegates work on a small engineering team.
Anthropic said it found no evidence that the group fielded an operational weapon, though the actors did test-fire a guided rocket. The launch failed, and within hours the actors returned to Claude to work out why, the Washington Post reported.
The report covers activity Anthropic said it disrupted between December 2025 and August 2026 across cyber operations, propaganda, surveillance and warfare. It also describes an Iran-linked actor compiling targeting handbooks on US Navy warships, Iranian state-aligned accounts running propaganda and domestic surveillance campaigns, and an influence operation tied to the United Arab Emirates that used around 300 fake social media accounts.
In a biological case, Anthropic said it blocked a request to help write a grant application for gain-of-function research on the chikungunya virus that could have made the pathogen more dangerous. It also flagged Russia-based developers using Claude to write software for autonomous kamikaze drone swarms, including terminal guidance and coordination between aircraft.
For founders building on frontier models, the report is a reminder that providers are actively policing misuse and that guardrails are not absolute, since the same coding tools that help ship products can be turned to other ends. Anthropic said it shared its findings with government and industry partners and urged other developers to watch for the same patterns.