OpenAI Disrupts Moonshot AI-Linked Distillation Campaign — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- OpenAI said on Wednesday it identified and disrupted a coordinated "adversarial distillation" campaign dating to the first week of July 2026, with a "core cluster" attributed to individuals associated with Beijing-based Moonshot AI; the company did not cite technical evidence, citing security reasons.
- The campaign began July 1, 2026 at low volume, spiked on July 24–25 to roughly 16,000 attempted requests from over 4,000 users, and showed related prompt-pattern activity across more than 15,000 users before being fully disrupted on July 28, 2026.
- OpenAI said the operators never broke its encryption or accessed stored user conversations but instead "manipulated model interactions" so protected reasoning could be reproduced in visible form at scale, in violation of its terms of service.
- In response, OpenAI deployed additional mitigations, banned the fraudulent accounts, closed a "pathway" that let someone with another user's encrypted reasoning replay and recover its contents, and added checks to detect streamed output that could expose reasoning.
- An August 2026 study by MATS Research, ELLIS Institute Tübingen, and Synk found an architectural flaw making encrypted reasoning traces "fully compatible and interchangeable" across sessions, users, and models within a provider's ecosystem, enabling scalable decryption jailbreaks.
- Moonshot AI now faces its second distillation accusation in weeks: Anthropic previously alleged it relayed customer requests to Claude instead of running them on its own Kimi model, activity tracked as GTG-16002.
Why it matters: OpenAI attributes the activity to individuals at a named Chinese AI lab without publicly naming the technical evidence, an unusual move that could escalate US–China tensions over model-IP theft while pulling Moonshot AI into a public dispute with two American frontier-model rivals within weeks.
Ask SkimNews
