OpenAI Says Chinese AI Labs Distilled GPT-6 Astra - Finimize

URL Source: https://news.google.com/rss/articles/CBMihAFBVV95cUxQQWxHOXBDdjR6RFhIXzJjWVFYazl1Mi16Q01VVU56NUVfVVIyYWZFY1A5Z3lsY0xSSE5PZ09HbzJwN3l0TUN2UC01T0pBSHZucW45S0swZFUxWW94TGd5NkJEeDRvazEzV3lVUDQ2NkNYSU5kRWtjcHpGNFFvZUg2RHpJMjU?oc=5

The Information says China’s DeepSeek and Moonshot AI trained recent models on outputs from OpenAI’s GPT-6 Astra, putting a harsh spotlight on “distillation” in the US-China AI race.

Distillation is when a developer uses an advanced model’s answers as training data for another model, a shortcut that can replicate capabilities without paying the full cost of original research. The Information reported that Moonshot’s Kimi K3.1 Pro and DeepSeek’s V4.1-Flash 1017 benefited from Astra outputs. OpenAI is treating that as more than academic: in a September 30 blog post, it said it linked a key part of a distillation campaign that began in July to ind..

to individuals associated with Moonshot AI. And it’s not just OpenAI: AI startup Anthropic said earlier this month that new protections for its Claude models cut data harvested in distillation-style attacks by about 90% to 99%, showing the industry is moving from building ever-bigger models to locking down access to them.

When labs treat distillation as an active threat, the fastest defense isn’t a lawsuit, it’s tightening the “pipes” into the model. That usually means more monitoring, rate limits (caps on how many requests you can make), and output restrictions designed to stop anyone from collecting large volumes of clean training data. Those controls can work – Anthropic’s 90%-99% estimate hints at the goal – but they also raise friction for legitimate high-volume users. So companies building AI features on top of OpenAI- or Anthropic-style application programming interfaces (APIs) may face stricter usage checks, more “prove you’re real” steps, and higher compliance overhead, which can slow feature rollouts and push up the cost of AI-powered tools that eventually gets passed on to customers.