← The Polylog AI Intelligence Brief
Morning Edition · Saturday, September 12, 2026Published at 2:23 AM EDT · New York
Anthropic Says DeepSeek and Moonshot Routed Customer Prompts to Claude Through Fake Accounts
The company reports 5,380 fraudulent Moonshot-linked accounts that relayed roughly 300,000 requests in ten days, and more than 12 million distillation attempts it attributes to DeepSeek over 14 days in July.
Anthropic's September threat intelligence report makes an accusation that goes beyond ordinary model distillation. The company says Moonshot AI and DeepSeek did not merely train on Claude's outputs. According to Anthropic, the two companies relayed live customer prompts to Claude and returned Claude's answers to users who believed they were talking to a Chinese model.
The figures Anthropic reports are specific. It attributes 5,380 fraudulent accounts to Moonshot, most of them presenting as based in Singapore and Japan to work around geographic restrictions, and says close to 300,000 customer requests were relayed over a single ten-day window, with most of that traffic sent to Opus. For DeepSeek, Anthropic says it observed more than 12 million distillation attempts over 14 days in July 2026. It also names operators affiliated with Alibaba as using Claude's outputs to help train Qwen models, according to CNBC's account.
The most consequential part of the report concerns data rather than model weights. If prompts were relayed as described, then whatever users typed reached Anthropic's servers. Anthropic says some of those exchanges contained sensitive material from individuals, multinational companies, and state-affiliated actors, including live credentials for a Russian government database and a query from a user it links to China's military about behavior tracked across surveillance cameras in Chengdu. The South China Morning Post reported the allegations as claims made by Anthropic, not as established fact.
Readers should weigh the evidentiary gap here. The account rests entirely on Anthropic's internal account telemetry and abuse investigation, which no outside party can inspect, and Anthropic is a direct commercial competitor of every lab it names. The accused labs have not published a detailed technical rebuttal. The claim is detailed and internally consistent. It is also unverified by anyone outside the company.
Veracity: Corroborated
78/100Part of a tracked trend
Chinese Open-Weight Models Emerge as the Non-US AI Stack
As Washington restricts foreign access to US frontier models, governments and enterprises cut off from American AI increasingly standardize on downloadable Chinese open-weight models, splitting the world into competing AI supply blocs rather than a single frontier.
Townsquare
Start a discussion in Townsquare.
More from this edition
- Clay Institute Declines to Call Navier-Stokes Solved After OpenAI's AI Proof Claim
- Anthropic Cuts Claude Fable 5.1 Cache Reads by 75 Percent
- Perplexity and Cognition Put GPT-6 Astra on Production Systems
- Anthropic Hands METR an Eight-Week Review After Claude Models Reached Live Systems
- Meta Claims Coding Leadership for Muse Spark 1.3, With Its Best Numbers From a Restricted Variant
- Sber Open-Sources a 432-Billion-Parameter Reasoning Model Under MIT License
- Sakana AI Splits Its Router Into a Cheap Tier and a Capability Tier
- Marigold V2 Turns a Diffusion Transformer Into a Single-Step Depth Predictor
Comments
0No comments yet.