Anthropic Says It Blocked Bioweapons Queries to Claude

Anthropic says it shut down accounts that tried to use Claude for work on dangerous pathogens and on weapons systems. The cases come from the company's first broad misuse report, described by German outlet heise online on September 11, 2026. Nobody outside the company has checked them: Anthropic is the sole source for what happened and how often.

One report, seven kinds of abuse

The document is titled "Detecting and countering misuse of AI" and runs to more than 150 pages, covering eight months of observation. It sorts incidents into seven areas: cyber operations, influence operations, surveillance, fraud, biological misuse, conventional weapons and model distillation. That is a far wider frame than the single-incident disclosures Anthropic has published so far.

The biology cases

One attempt used Claude to help draft a funding application for gain-of-function research on a virus. In another, thousands of messages about influenza viruses arrived through a VPN and an anonymous email address. Der Standard adds a plan to study an avian flu variant that could jump to mammals, and reports that requests from geographically high-risk regions were restricted.

The hard part is dual use. From the outside, a question about a pathogen looks much the same whether the goal is a vaccine or a weapon, which is why the report's judgment calls matter as much as its counts.

Missiles, torpedoes, drone swarms

Under conventional weapons, the report describes actors from northern Yemen working on missiles and Chinese efforts around anti-torpedo technology. It also cites what heise online renders as suspected freelance actors from Russia developing autonomous kamikaze drone swarms. Attribution in all three cases is Anthropic's own assessment, not a finding by any authority.

Rivals routing queries into Claude

The distillation section carries the most concrete figures. Anthropic counted 23 million interactions traced to Moonshot (Kimi) between May and July, and 12 million through Deepseek inside 14 days. The allegation is that competitors quietly passed user queries to Claude rather than answering them with their own models.

What is not settled

A third German-language report on the same document could not be opened for this article, so none of its details are used here. Two questions stay open: how Anthropic ties individual accounts to specific actors, and what "blocked" means in practice — a ban, a refusal filter, or a referral to authorities. Until the report is examined independently, it remains one company's account of its own enforcement.