Anthropic Accuses Chinese Giants of 190 Million Attacks

Key Takeaways

1Anthropic accuses several Chinese laboratories of conducting nearly 190 million distillation attacks on Claude between May and July

2The company details massive query hijacking, involving Alibaba, Moonshot, DeepSeek, Zhipu, and Xiaomi

3Sensitive data related to Chinese and Russian governments and military may have been exposed

4Anthropic announces technical measures and targeted identity verifications

Why it matters : These accusations illustrate the extent of the risks associated with the hijacking of advanced models and the technical response from a major player in the industry.

Anthropic claims to have faced nearly 190 million distillation attacks between May and July, carried out by several Chinese laboratories. The company mentions massive redirections of requests to Claude, exposure of sensitive data, and announces technical safeguards and targeted identity checks.

Exposed Sensitive Data and Record Volume Attributed to Alibaba

Anthropic reports that requests intended for Chinese models but redirected to Claude have exposed sensitive information related to Chinese and Russian authorities. The company cites surveillance videos uploaded by a user affiliated with the People's Liberation Army and elements from a Russian government database submitted by a Russian military contractor. Meanwhile, Anthropic attributes the largest attack it has measured to Alibaba, with over 151 million exchanges between May and July, conducted by 3,500 fraudulent accounts. According to Anthropic, these accounts led Claude to detail its reasoning processes, which were then exploited to train Qwen models.

Who is Targeted and How Requests Were Redirected

Anthropic claims that DeepSeek and Moonshot AI redirected user requests to Claude instead of their own systems. The company states it recorded 23 million redirections for Moonshot between May and July, and 12.1 million for DeepSeek over 14 days in July. Anthropic also accuses Z.ai, linked to the Ox Alpha model, of a similar attack to that attributed to Alibaba, and asserts that Xiaomi recorded conversations from its MiMo models to inject them into Claude to generate training data. At the heart of these practices, distillation involves relying on the outputs of a more advanced model to improve a simpler model. In total, Anthropic estimates at least 35 million request redirections carried out during the summer by major Chinese laboratories.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

What Anthropic is Changing in Claude and Access to the Service

To prevent further campaigns, Anthropic states it has implemented detection mechanisms and bans on suspicious activities. The company explains it has adjusted Claude to produce less detailed reasoning transcripts, deemed less exploitable for third-party training. It also plans to require identity verification for users appearing to operate from China, Russia, or Iran, countries where the service is not available.

Public Background and Quantitative Assessment of the Period

Between May and July, Anthropic reports a total close to 190 million distillation attacks attributed to Alibaba, Moonshot, DeepSeek, Zhipu, and Xiaomi. The company describes a bombardment of requests during the summer and implicates several Chinese actors. It recalls having previously publicly called out these laboratories. In June, Sarah Heck, head of policy at Anthropic, indicated in a letter to lawmakers that Alibaba had illicitly conducted 28.8 million exchanges with Claude between April 22 and June 5, calling for legislative intervention.