Anthropic: Alibaba Ran 29 Million Fake Queries to Clone Claude
Anthropic: Alibaba Ran 29 Million Fake Queries to Clone Claude
Publish Date: 2026-06-26 18:14:00
Source Domain: www.pymnts.com
-
Model Distillation Costs: Training a frontier artificial intelligence (AI) model is prohibitively expensive, while reproducing its behavior through model distillation can significantly reduce costs.
-
Anthropic’s Accusations: Anthropic accuses Alibaba and its AI lab of conducting the most extensive known distillation campaign against its models. The campaign allegedly generated over 28.8 million interactions using around 25,000 fraudulent accounts between April 22 and June 5.
-
Scale of the Campaign: The scale of this campaign is significantly larger than previous instances, where three Chinese AI labs generated more than 16 million interactions through 24,000 fraudulent accounts in February.
-
Detection Challenges: Distillation is difficult to detect because it mimics legitimate queries. Detection relies on recognizing patterns such as large volumes, repetitive prompts, and coordinated account activity.
-
Safety Concerns: Distillation without authorization does not inherit the safety guardrails of the original model, risking harmful capabilities being transferred to the competing model.
-
Legal and Policy Push: Anthropic is advocating for legislative action to make model theft illegal, supported by bipartisan efforts and the White House Office of Science and Technology Policy to sanction entities conducting such campaigns.
-
Economic Implications: If adverse distillation becomes commonplace, it could force AI labs to spend more on security and access control measures, potentially impacting the business model of AI as a service.
