Anthropic says Alibaba, Moonshot AI and DeepSeek used large-scale distillation campaigns involving nearly 200 million Claude exchanges to extract advanced capabilities for competing AI models. (Source: Image by RR)

Distillation Efforts Targeted Reasoning, Coding and Agentic Capabilities

Anthropic says China-based AI companies have conducted increasingly sophisticated campaigns designed to extract capabilities from Claude and use them to improve competing models. In a new threat intelligence report, the company attributes five separate campaigns involving nearly 200 million exchanges to distillation efforts targeting some of Claude’s most valuable abilities, including agentic behavior and tool use, coding, data analysis, logical reasoning and access to detailed reasoning traces. Anthropic says the campaigns have grown larger and more aggressive as competition among frontier AI developers has intensified.

Model distillation, as noted in an article at techcrunch.com, is a legitimate technique when performed with authorization, allowing a smaller model to learn from outputs generated by a more capable system. Anthropic alleges these campaigns instead circumvented its restrictions and attempted to extract Claude’s hidden reasoning processes at industrial scale. Claude normally exposes only summarized descriptions of its reasoning rather than its internal chain of thought, but attackers reportedly discovered prompts capable of eliciting more detailed traces, including one that disguised the extraction as a request to translate the model’s working memory into katakana-only Japanese.

The largest campaign was attributed by Anthropic to Alibaba and involved approximately 151 million exchanges between May and July 2026, peaking at nearly 3 million interactions per day across roughly 3,500 accounts. Anthropic says the accounts shared a fixed extraction prompt, leading investigators to conclude they were part of a coordinated effort to generate training material for Alibaba’s Qwen models. The company also attributed separate campaigns to Moonshot AI and DeepSeek as part of the broader effort to harvest capabilities from Claude.

Anthropic says the Moonshot campaign also raised national-security concerns because some requests appeared to be routed from China’s military. During one 10-day period, nearly 300,000 requests reportedly reached Claude through approximately 5,000 accounts, primarily targeting its Opus model, including a request to evaluate surveillance footage for abnormal behavior. The findings add a new dimension to the global AI competition, where access to a rival’s advanced model can potentially provide not only useful answers, but training material that accelerates development of competing systems.

read more at techcrunch.com