Anthropic report on Claude misuse
This report shows how Chinese AI companies are using Claude to train their own models, which could impact the quality and security of their products.
Anthropic's September threat intelligence report details how Russian spies, Chinese hackers, scammers, and bots have used Claude for espionage, 0-day hunting, mass dating scams, fake news, and building surveillance systems.
It also found that Chinese companies like Alibaba, DeepSeek, and Zhipu are stealing Claude's answers to train their own models.
For example, Alibaba's Qwen 3.5/3.
6/3.7, Moonshot's Kimi, and DeepSeek's models all use this method.
Xiaomi's MiMo also uses Claude to process user data for training, and SenseTime buys Claude data from third parties.
MiniMax operates a shell company to collect user data from US models.
太有意思啦,Anthropic 9月威胁情报报告: 过去8个月,俄间谍、中国黑客、骗子、水军都在用 Claude 搞间谍攻击、挖0day、批量约会诈骗、造假新闻、搭监控系统,还有人拿它写导弹制导软件、做双用途生物课题。 而阿里、DeepSeek、月之暗面、小米、智谱、商汤、MiniMax 都被点名,批量偷 Claude 答案去训自家模型。月之暗面和 DeepSeek 更狠,用户以为在用 Kimi / DeepSeek,后台其实在跑 Claude。(赶紧把滑祖器拨到梁圣去) 几家大厂怎么“抄Claude作业”: - 阿里:体量最大。高峰一天抽近300万条,3500多个假号,整段归因超1.51亿,专扒 Opus 4.6/4.7 的思考过程去喂 Qwen 3.5/3.6/3.7。 - 月之暗面:用户请求悄悄转给 Claude,再把回复当成 Kimi 吐回去。10天转了近30万条,整段归因超2300万。还绕过 thinking signature,把思考过程抠出来训自己的模型。 - DeepSeek:同一套。14天超1210万条。专门把走 Claude Code / OpenCode 的用户请求转去 Opus。 - 智谱:把偷来的 Claude 思考过程再喂回 Claude洗干净,好拿去训自家 GLM。10天273个假号专抽 Opus 4.8,光清洗就77万条,17天总共归因340万条,还拿 Claude 当裁判给训练数据打分。GLM 5.3 发布前,用公开漏洞集出 CTF 题,去套另一家美国顶模的黑客能力。Fable 防护太硬啃不动,转头打防护更弱的 Opus 4.6,主要拿它给那家模型的答案打分。 - 小米MiMo:把自家用户跟 MiMo 的聊天、写代码记录回放给 Claude 做训练数据,并不拿去给用户看。20天40多万条,走了1500多个代理号。时间点刚好卡在 MiMo-V2-Pro 免费试用结束时,疑似靠免费吸引全球开发者来刷量,好批量收割去蒸馏。 - 商汤SenseTime:自己不直接偷,从第三方数据商买别人从中转站扒下来的 Claude 对话记录,还让 Claude 帮自己写蒸馏流水线。 - MiniMax:开了个壳公司做代理,只卖 Anthropic 和 OpenAI 的访问,连自家模型都不挂。报告推断:挂这个代理就是为了把用户跟美国模型的对话存下来训模型。 都是一帮神人啊,哈哈哈哈,我好像有点理解 Anthropic 了,身上全是蚂蟥。 Anthropic @AnthropicAI We're publishing our most detailed threat intelligence report to date. It covers how people tried to misuse Claude—for cyberattacks, influence operations, surveillance, biology, and building weapons—and how we found and stopped them. We disrupted every operation in the report, and used the lessons from them to strengthen our safeguards. Where appropriate, we also shared what we found with authorities and other AI companies. These cases are not typical: we’re highlighting some of the most sophisticated misuse we’ve seen. But they’re especially important to discuss, because they show us where AI misuse is headed, where our safeguards work, and where they need to improve. We’re publishing this report so others can spot the same activity on their own platforms, and so we can give the public a clearer view of how emerging threats develop. Read the report: anthropic.com/threat-intelli… 🔗 View Quoted Tweet📊 16 ⚡
This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Claude, DeepSeek, Moonshot AI, Qwen.
Source attribution: curated from Geek (@geekbb). Cross-source corroboration: 11 independent sources have covered this development (including shao__meng, techcrunch, Decoder), confirming industry-wide signal strength.