Anthropic has published a new threat intelligence report describing a wave of distillation attempts aimed at extracting model behavior from its Claude family. The company says the activity has grown more sophisticated as global competition in AI intensifies.
According to the report, the campaigns sought to capture capabilities such as agentic workflows, tool use, coding, data analysis, and logical reasoning. Anthropic says the effort involved nearly 200 million exchanges across five separate campaigns.
Distillation is a technique that can be used to train smaller models by learning from the outputs of larger systems. Anthropic says some actors found ways to prompt Claude into revealing its internal reasoning traces, including by disguising requests as translation tasks.
The largest campaign described in the report was linked to Alibaba, with 151 million exchanges recorded between May and July 2026. Anthropic says the activity spread across 3,500 accounts and appeared designed to generate training data for the Qwen model family.
The report also points to a campaign associated with Moonshot AI, maker of Kimi. Anthropic says some requests were routed through a network of 5,000 accounts and focused heavily on the Opus model, including analysis of surveillance footage for behavioral assessment.
Anthropic notes that similar pressure on frontier AI systems is likely to shape how model security, access controls, and evaluation methods evolve in the years ahead.
