The most important theft in AI doesn't involve a server room. It involves a credit card and an API key.
Chinese labs DeepSeek and Moonshot AI "relied heavily on outputs from OpenAI's powerful GPT-6 Astra model to train their latest models," The Information reported on Wednesday, citing people with direct knowledge. The models in question are Moonshot's Kimi K3.1 Pro, released last week, and DeepSeek's upgraded V4.1-Flash 1017, launched October 17. The technique is distillation: query a stronger model at scale, collect its answers, and train a cheaper model on them. OpenAI, DeepSeek, and Moonshot AI all declined to comment.
The economics explain why everyone does it. AI Weekly notes that on a third-party browser-operation benchmark, V4.1-Flash 1017 moved from 70.9 to 75.0, closing on GPT-6 Astra's 77.9 — while costing $0.22 per task against Astra's $10.77, a nearly 49-fold gap. The student nearly matches the teacher at two cents on the dollar. And there's a sting in the tail: Moonshot's own numbers put Kimi K3.1 Pro at 65.7% on Terminal-Bench 4.0 — ahead of GPT-6 Astra's 58.2%. The distillate now outscores the source, at least on one test.
This is the hole in Washington's entire AI strategy. The government has spent the year restricting the hardware — banning DeepSeek from federal agencies, scrutinizing chip exports — while frontier intelligence diffuses through commercial API endpoints that anyone can query. Every prompt answered is a potential training example for a competitor. The labs know it: OpenAI said in a September 30 post that it had linked a distillation campaign begun in July to individuals associated with Moonshot AI, and Anthropic said earlier this month that new defenses cut harvested data by 90 to 99 percent. Dealroom reports that Chinese labs simply shifted targets from Claude to Astra once Anthropic hardened its defenses — and cites a September 8 advisory from the NSA, FBI, and CISA alleging six Chinese labs have targeted Claude, GPT, Gemini, and Grok since at least late 2024. Beijing's commerce ministry called those claims "groundless and legally unsound."
So the distillation wars have become a shell game: harden one endpoint, and the queries move to another. That's why this report matters more than a benchmark table. Anthropic is now asking courts to block risky models, OpenAI wants mandated kill switches — but neither addresses the leak. Until someone figures out how to sell intelligence through an API without giving it away, the frontier will keep getting photocopied. Watch whether OpenAI sues, whether the API terms get teeth, and whether Moonshot's promised November 4 open-weight release makes K3.1 Pro's distilled gains permanent — downloadable by anyone, forever.
— via reuters