Anthropic says Alibaba, Moonshot AI, and DeepSeek tried to copy Claude

In short: Anthropic says it detected repeated attempts to copy parts of its Claude AI model by collecting large numbers of answers from it.

What's going on

Anthropic published a new threat intelligence report saying it observed “distillation attacks” aimed at Claude. Distillation is a way to make a smaller AI act more like a bigger one by feeding it lots of the bigger model’s answers (like copying a chef by writing down thousands of their recipes).

According to Anthropic, these efforts tried to get Claude to reveal more of its step by step thinking than it normally shows. Anthropic says Claude usually provides only “summarized thinking,” which is a short explanation instead of the full internal notes. The report says attackers found prompts that could trick the system into showing more detailed thinking traces, including one that framed the request as a translation into katakana-only Japanese.

Anthropic says it saw nearly 200 million “exchanges,” meaning individual back and forth interactions, linked to five campaigns. It attributed the largest campaign to Alibaba, saying it involved about 151 million exchanges from May to July 2026 and spread across roughly 3,500 accounts, with a peak near three million exchanges per day. Anthropic also described a campaign it linked to Moonshot AI, saying some requests appeared to be routed from the Chinese military, including a request to review surveillance footage for “abnormal” behavior.

What to watch

If more companies try to gather training data this way, AI providers may tighten access, add more checks, or limit what their systems explain. That could affect everyday users by making AI tools more restrictive or less transparent about how they reached an answer.

Source: TechCrunch AI