Anthropic claims that multiple Chinese AI companies illegally invoked Claude for training models
CNBC
13h ago
Ai Focus
Anthropic claims that Alibaba and the 'Dark Side of the Moon' once called upon Claude on a large scale to generate training models, with some requests also involving sensitive information.
Helpful
No.Help

Anthropic released a threat intelligence report stating that between December 2025 and August 2026, the company detected and blocked multiple large-scale abnormal calls targeting Claude, involving Chinese AI companies such as Alibaba, Moon's Dark Side, and DeepSeek. The report classified these incidents as unauthorized model distillation, which refers to using the output of a more powerful model to train another model.

Alibaba is accused of having the largest scale of...

According to Anthropic, the entity associated with Alibaba once used Claude to output a series of models called "Tongyi Qianwen" Qwen. This is also considered the largest distillation operation detected by the company to date.

According to the report, this activity occurred from May to July, involving over 151 million interactions with Claude in total, with a peak of nearly 3 million interactions per day. These requests are said to have come from more than 3,500 fraudulent accounts. Anthropic also stated that the same entity used Claude for reinforcement learning and model architecture research as well.

The dark side of the moon is said to have forwarded a request for Kimi.

Anthropic Also mentioned Beijing. AI The dark side of the company. According to the report, the 'dark side of the moon' once forwarded some user requests originally intended for Kimi to Claude, and then displayed the responses from Claude to the users, without the users knowing that the actual responses came from Anthropic.

The report states that within a 10-day period, the 'Dark Side of the Moon' forwarded nearly 300,000 customer requests to Anthropic, with the vast majority of these requests being routed to the Claude Opus model. These requests were initiated from 5,380 accounts identified as fraudulent, and most of these related accounts were located in Singapore and Japan.

Anthropic further stated that the 'Dark Side of the Moon' has preserved at least some of the conversation records and extracted the reasoning text from Claude to use as training data for its own model. During the period from May to July, there were over 23 million interactions attributed to the 'Dark Side of the Moon'.

DeepSeek is also included in the report.

According to Anthropic, DeepSeek also adopted a similar approach to the "dark side of the moon," transferring some interactions to Claude without notifying the users. The company stated that within just 14 days of July 2026, they observed over 12 million distillation attacks that could be attributed to DeepSeek.

Anthropic also indicates that some requests forwarded to Claude contain sensitive information from individual users, large multinational corporations, and entities associated with countries. The report suggests that such practices may be inconsistent with privacy laws and the service terms of the relevant companies.

Additional information:This report covers various types of abusive behaviors that Anthropic interrupted between December 2025 and August 2026, including not only model distillation but also cyber operations, influence operations, monitoring, fraud, biological abuse, and conventional weapons development. As of the time of publication on CNBC, Alibaba, Dark Side of the Moon, DeepSeek, Xiaomi, and Anthropic have not immediately responded to requests for comments.

Tip
$0
Like
0
Save
0
Views 27
CoinMeta reminds readers to view blockchain rationally, stay aware of risks, and beware of virtual token issuance and speculation. All content on this site represents market information or related viewpoints only and does not constitute any form of investment advice. If you find sensitive content, please click“Report”,and we will handle it promptly。
Submit
Comment 0
Hot
Latest
No comments yet. Be the first!
Related
Foreign media: Reddit User growth drives a 6% increase in stock price
Foreign media reports that Reddit's stock price rose by about 6% due to increased user growth data, but it still faces significant resistance in the range of $159 to $161.
The Cryptonomist
·2026-09-11 22:14:44
8
web3: Foreign media: OKB rises to a key resistance zone, $120 becomes the focus for short-term trading
OKB breaks through the consolidation range and approaches the $120 mark, while foreign media reports that X Layer TVL has risen to approximately $232 million, providing fundamental support for its trend.
CoinPedia
·2026-09-11 21:45:33
16
web3: Trezor claims email service provider has been hacked, users face phishing risks again
Trezor claims that third-party email service providers have been hacked, and users' email addresses are now at risk of a new round of phishing attacks after being exposed.
TechCrunch
·2026-09-11 21:45:30
14
web3 : The XRPL lending protocol is to be launched, but the supporter rate among verifiers still does not meet the standard.
XRPL The native lending protocol is scheduled to be launched with version 3.4.0, v1.1. However, the supporter rate of the relevant amendments still falls below the 80% activation threshold.
Coinpaper
·2026-09-11 20:11:22
22
The United States relaxes some travel restrictions for air travel, and visits at boarding gates are resumed
The United States begins to relax some of the aviation restrictions implemented after September 11, while TSA also emphasizes that AI, drones, and cybersecurity are becoming new sources of risk.
CNBC
·2026-09-11 19:18:14
26
View More