# Anthropic 2026 年 9 月滥用报告：拦截危险生物请求，指认 7 家中国实验室未经授权蒸馏 Claude

- 来源：Rohan Paul (@rohanpaul_ai)
- 发布时间：2026-09-11 09:03
- AIHOT 分数：68
- AIHOT 链接：https://aihot.news/items/cmtw9hg9k08ufrolk1v06at08
- 原文链接：https://x.com/rohanpaul_ai/status/2098215738939220371

## AI 摘要

Anthropic 发布 2026 年 9 月滥用检测报告，称已拦截旨在增强病毒危险性的请求，并描述了 5 起生物滥用案例，涉及禽流感研究和一份来自军方研究机构的基孔肯雅热资助申请书。

## 正文

Anthropic latest "misuse of AI" report claims some really sensational stuff.

Says that Anthropic blocked requests to make viruses more dangerous and uncovered Chinese campaigns copying Claude’s capabilities.

The report describes 5 biological cases, including bird flu research and a chikungunya grant proposal involving a military research institute.

Account bans did not end access: a reseller serving the chikungunya researchers returned within days, and their research continued.

Anthropic previously assessed older Claude models as too limited to meaningfully help experts conduct dangerous biological research. But, it cannot give the same assurance for newer models, so it now restricts a wider range of biological requests.

Anthropic separately identified unauthorized distillation campaigns involving 7 Chinese labs.

It says DeepSeek and Moonshot also forwarded customer requests to Claude and presented its replies as their own models’ output.

Those transfers exposed sensitive company information and, in some cases, live credentials to Anthropic, apparently without customers’ knowledge.

Customers cannot control where sensitive information goes if providers silently substitute the model processing it.