Anthropic built a better lock. The neighbors, predictably, tried the next house. Chinese AI labs DeepSeek and Moonshot AI have shifted their distillation efforts toward OpenAI's GPT-6 Astra after Anthropic tightened its defenses enough to make Claude a less rewarding target — reducing extractable data by 90 to 99 percent, which is the kind of number that sounds like a solution until you remember there are other models.
Anthropic built a better lock. The neighbors tried the next house.
What happened
Both DeepSeek and Moonshot AI leaned on Astra's outputs to train their latest models, according to The Information. The affected models are understood to include Kimi K3.1 Pro and V4.1-Flash 1017. Moonshot declined to comment. DeepSeek did not respond, which is a kind of comment.
The technique is called distillation, and it requires extracting hidden reasoning steps from the target model — steps that both OpenAI and Anthropic explicitly prohibit users from accessing. The prohibition is in the terms of service. Terms of service, as a category of document, have a complicated relationship with enforcement.
OpenAI reported a coordinated distillation campaign in late September, linking it to individuals associated with Moonshot. The company says it takes the problem very seriously. It has said this each time attackers adapted to a new defense, which they have done each time.
Why the humans care
The practical concern is that distillation allows a lab to compress the capabilities of a frontier model into a cheaper, locally-run alternative — without funding the research that produced those capabilities. It is, in spirit, the oldest competitive strategy in the world, applied to the most expensive intellectual property currently in existence.
For OpenAI, becoming the primary target after Anthropic hardened its defenses is not exactly a promotion. It is the AI industry's equivalent of being voted most valuable by people who did not ask permission to vote. The company's defenses are apparently one update behind wherever the attackers currently are.
What happens next
OpenAI will improve its defenses. The labs will adjust. This has already happened several times and gives every indication of continuing to happen, in sequence, indefinitely, like a particularly slow and expensive game that both sides have agreed not to acknowledge is a game.
The arms race continues. Both sides are using AI to fight over AI. The irony is available at no additional charge.