The Story: Anthropic disclosed that it detected what it describes as coordinated distillation campaigns involving over 16 million exchanges across 24,000 fake accounts, allegedly tied to rival AI labs attempting to extract and replicate Claude’s capabilities.

What Anthropic Claims Happened
According to the company, the activity involved large-scale conversational interactions designed to systematically capture Claude’s outputs. Anthropic characterizes this as “distillation via conversation,” where a weaker system is trained on the outputs of a stronger frontier model.

Scale of Activity
Anthropic stated that MiniMax allegedly ran the largest campaign, with more than 13 million exchanges before being detected mid-operation. The company claims it observed rapid shifts in activity patterns within 24 hours of intervention.

DeepSeek was described as prompting Claude to provide detailed step-by-step reasoning and to rewrite politically sensitive queries — potentially generating structured training data for reasoning improvement and policy alignment systems.

Industry & Policy Implications
Anthropic noted that OpenAI had previously raised similar concerns with U.S. lawmakers, and is now calling for broader coordination between AI companies and governments to address large-scale model extraction attempts.

The company framed the issue as a security and economic risk to frontier model developers, arguing that distillation at scale undermines incentives for costly research and infrastructure investments.

Why It Matters: As global AI competition intensifies, distillation and model extraction are emerging as central battlegrounds. Frontier labs invest billions in training and infrastructure; if competitors can replicate capabilities via systematic querying, the economics of frontier development could shift dramatically. At the same time, the broader AI industry faces scrutiny over its own data practices — making the debate over “who trained on whom” increasingly complex.

Source: Anthropic Official Disclosure