Anthropic report: Claude misused for missiles, spy malware, 151 million exchanges mined

Anthropic's report breaks misuse into seven categories and says it documents novel cases rather than the typical kind. The affected models were primarily Haiku, Sonnet, and Opus, while Fable and Mythos appeared in only a single distillation case — suggesting that older, widely deployed models remain the primary vector even as newer ones carry stricter safeguards.

The report's cyber chapter argues that sophisticated attacks no longer require sophisticated attackers. The techniques themselves are familiar — stolen credentials, unpatched devices, SQL injection, phishing — but reconnaissance, exploitation, and tool-building now run in parallel at machine speed. A Russian-speaking actor tracked as GTG-20006 ran AI agents that checked whether malware was being flagged, rewrote and recompiled the code when it was, and slipped past detection again; more than 20 organizations were targeted, and the actor stole a complete proprietary SDK for a drone vision system.

The distillation findings raise a different set of questions. Anthropic says the illegitimate version is industrial-scale and covert, enabled by fake accounts and 'transfer stations.' But the strangest cases involve labs relaying their own customers' requests — Moonshot AI (GTG-16002) relayed nearly 300,000 customer requests over ten days while users believed they were using a Kimi model, and DeepSeek (GTG-16001) routed selected users to Claude Opus, over 12.1 million exchanges in 14 days. Among those, Anthropic found a user likely tied to the People's Liberation Army analyzing CCTV footage from hundreds of cameras in Chengdu, and an operator with live credentials for a database linked to the Russian Ministry of Defense. Whether stricter API controls can prevent this kind of relay — where the misuse happens through intermediaries rather than direct accounts — is the open question the report leaves unresolved.