Anthropic claims DeepSeek, Moonshot AI and Minimax are sweeping its data

AI is improving rapidly, but the drama behind the scenes is unfolding twice as fast. In the latest scenario, Anthropic, the company behind Claude, alleged that DeepSeek, Moonshot AI, and MiniMax are sweeping its data without safety training practices.

The AI company says these groups set up over 24,000 phony profiles and sent more than 16 million queries. One team zeroed in on deep reasoning skills, another grabbed coding and agent-style tasks, while the third pulled tool-use patterns right before they launched their own models. When Anthropic rolled out fresh updates, the attackers switched strategy in less than a day.

Distillation is the key to this discussion. Teams feed answers from a strong model into a smaller one, so the smaller version picks up those same skills fast and with less cost. People do this the right way all the time to shrink models for phones or simple gadgets.

The trouble starts when the AI team does it on the sly. It skips every safety check the original team built in. Attackers craft the same prompts again and again to pull out step-by-step thinking, tool calls, and sharp analysis. They spread the traffic across synced accounts so they dodge limits and blend in with normal users.

Anthropic found this movement by watching API traffic for odd patterns and using tools that flag behavior normal folks never show. They now track coordinated activity and strange details in the data. The AI firm has shared warning signals with other labs, cloud providers, and officials. They also filled cracks for research and startup access while adding protections at every layer. This includes the product down to the model core. These steps slow the theft without slowing real work.

However, this allegation has been discussed online discussion. People point out that open versions of the copied models could spread the risks even faster. The timing lines up with bigger conversations about controlling heavy compute power and sales limits, tying straight into wider security questions.

This stuff matters because secret distillation rips away the careful barriers that keep powerful AI out of dangerous spots like bioweapons or cyber ops. It lets others leap years ahead and skip years of safety work. Still, Anthropic’s fast response shows teams can push back when they catch problems early.