Anthropic’s Threat Report Makes Distillation an Industrial Story
Anthropic’s September threat report alleges large-scale distillation, including more than 151 million Alibaba-linked exchanges, and a deceptive dating-app network with over 4,700 AI personas. It describes five concerning biology cases while explicitly declining to infer intent to harm.
Anthropic’s September threat report describes alleged misuse at large scale. Its attribution and incident counts are the company’s findings, not independently adjudicated facts.
Anthropic (September 10) published “Detecting and countering misuse of AI: September 2026.” Window: December 2025–August 2026. Seven harm areas: cyber, influence, surveillance, scams, biological misuse, conventional weapons, illicit distillation. Anthropic’s reported cases include: a Yemen-linked cell using Claude “in place of human software engineers” for missile-guidance work; an actor whose attribution Anthropic says is consistent with Midnight Blizzard, using automated workflows; five biology cases with account bans after investigation (including chikungunya GoF grant language and mammal-adapted HPAI) — Anthropic does not assert the scientists intended harm; China dating-app studio GTG-15001: >4,700 AI personas, 25,000 people in two weeks; Alibaba GTG-16005: >151 million CoT exchanges vs Opus 4.6/4.7 (May–Jul); Moonshot 23 million; DeepSeek serving Claude instead of own models. Anthropic says it disrupted the reported activity and shared intelligence with authorities and partners where appropriate.
Limits#
- The named organizations’ responses and any independent investigations could change how these allegations are assessed.
- Account bans and safeguards do not mean every concerning request was blocked before an answer; Anthropic does not claim the scientists intended harm.
Correction (October 14, 2026): Strengthened attribution for company allegations and clarified that disrupted biology cases were not all blocked before model assistance; preserved the explicit no-harmful-intent caveat.