Anthropic says Claude helped scale surveillance and weapons work

Anthropic says it shut down Claude accounts used by suspected state-linked actors, armed groups, and others for surveillance, cyber activity, weapons-related engineering, and potentially dangerous biological research. Its report covers cases it says it disrupted between December 2025 and August 2026, and describes AI less as a source of new expertise than as an always-on assistant for research, coding, dividing tasks, and checking work across many parallel sessions. The company also alleges that China-based AI firms used large networks of fake accounts to collect Claude’s answers for training rival models—a shortcut known as distillation. These are selected cases rather than a measure of everyday use, but they show why AI safety is becoming as much about monitoring access and misuse as about what a model will answer in a single chat.