nthropic announced that it has identified three AI laboratories DeepSeek, Moonshot AI and MiniMax for launching an "industrial-scale" distilled capability extraction operation against Claude: more than 16 million interactions with Claude were generated through about 24,000 fraudulent accounts to train and improve its own models, and were suspected of violating terms of service and regional access restrictions.
The announcement pointed out that "distillation" itself is a common training method, but when competitors use fraudulent accounts, proxy services and batch high-repetitive prompts to centrally extract differentiated capabilities such as reasoning, tool use and coding in a short period of time, it will weaken security guardrails and bring security risks. Among the three operations, Moonshot had about 3.4 million interactions, MiniMax had about 13 million interactions, and DeepSeek was small but included strategies such as inducing output of "step-by-step inference trajectories".
Anthropic said it has strengthened the verification of abuse-prone channels such as detection and behavioral fingerprinting, improved education and research, and shared technical indicators with other institutions. Countermeasures at the product, API and model levels are also developed to reduce the effectiveness of outputs being used for illegal distillation. The announcement also mentioned that such attacks may affect policy discussions on model exports and computing power restrictions, but the relevant external impact still depends on the follow-up actions of all parties.
FAQs
Q: Who is the subject of this incident?
A: The announcement was made by Anthropic and involved its model Claude and the named DeepSeek, Moonshot AI, and MiniMax.
Q: How is Distillation Attack different from normal distillation?
A: Normal distillation is mostly internal compression and cost reduction in the same institution; Distillation attacks refer to the extraction of competitor capabilities through fraudulent accounts and large-scale prompts.
Q: What is the scale data disclosed?
A: Approximately 24,000 fraudulent accounts and more than 16 million interactions with Claude; Among them, Moonshot has been about 3.4 million times and MiniMax has about 13 million times.
Q: What capabilities are mainly extracted from this type of attack?
A: The announcement mentions that the focus includes reasoning ability, tool use, coding and agent tasks, etc.
Q: What risks do ordinary users need to be aware of?
A: If the model proliferation without guardrails may amplify the risk of abuse; Users should also be wary of services such as agent resale and abnormally low-price "proxy access".
