OpenAI, Anthropic, and Google — typically fierce competitors — announced a joint security initiative in early April 2026 to combat what they describe as systematic model theft by Chinese AI companies. The three labs are sharing intelligence through the Frontier Model Forum, a body they co-founded, and have named three specific firms: DeepSeek, Moonshot AI, and MiniMax. Anthropic claims these three companies collectively generated over 16 million exchanges with Claude through approximately 24,000 fraudulent accounts, using the conversations to train their own models in a practice known as adversarial distillation.
Adversarial distillation is a technique where a competitor queries a closed model at scale and uses the responses as training data, effectively teaching a new model to mimic the original without paying for the underlying research. It is technically difficult to prevent entirely because it looks like normal API usage until patterns emerge at large scale. The three labs say they have now built shared detection tools and will report suspected distillation attempts to each other in near-real-time.
The move reflects a broader shift in how the AI industry is thinking about intellectual property. Copyright law has not kept up with AI training practices, and courts have so far given mixed signals on whether scraping-for-training is permissible. By framing the problem as a security and national-security issue rather than a copyright one, the US labs are signaling that they may push for new legislation or international agreements rather than waiting for courts to catch up.
For students, this story illustrates a tension that will shape AI careers: the field has thrived on open sharing of research, but as models become more commercially valuable, that openness is being tested. Understanding the legal, ethical, and technical dimensions of how AI systems are trained — including where training data comes from — will be an important part of working responsibly in this industry.