According to CNBC, OpenAI identified a sustained effort beginning in July that escalated to 16,000 requests from over 4,000 users within two days, with related activity spanning more than 15,000 accounts. The attackers targeted protected model reasoning —the internal logic layer that powers frontier AI systems. They weren't trying to breach encryption, databases, or stored conversations. OpenAI traced a core cluster of the activity to Moonshot AI, the Chinese startup behind the Kimi chatbot.
This signals a shift in how rivals compete for AI advantage. Competitors are now after the expensive reasoning layer that makes frontier systems work, not just model weights or user data. If they can extract that through mass queries—using the product exactly as designed until the pattern reveals the strategy—the competitive moat around leading labs erodes fast. For geopolitics, this complicates the assumption that compute is the only choke point.
Most AI security talk still sounds like old-school hacking. Stolen credentials. Zero-days. Data exfiltration. This looks different: industrial extraction through the front door, at scale, using the product as offered until the volume and pattern give it away. That matters because Washington's export controls and China's catch-up race both assume chips are the bottleneck. They might be. But know-how is another one, and OpenAI just said someone came shopping for it in bulk. The AI race may be shifting from hardware to behavior.
Filed to the World desk · 6 days ago