OpenAI (explore this topic) Accuses Moonshot AI Users of Extracting Its AI Reasoning
OpenAI has revealed that users linked to Moonshot AI, the Chinese developer of Kimi, attempted a coordinated effort to extract the hidden reasoning behind its AI models. This revelation was shared in a blog post on Wednesday.
The activity began on July 1st at a low volume and peaked on July 24th and 25th, with 16,000 requests from over 4,000 users. OpenAI identified related activity across more than 15,000 users and successfully shut down the campaign by July 28th. While they couldn't attribute every operation to a single actor, they identified a "core cluster" linked to Moonshot.
“This activity is consistent with adversarial distillation: the systematic and unauthorized use of one model’s outputs or reasoning to help train, reproduce, or improve another model,” OpenAI stated.
How the Extraction Was Carried Out
OpenAI encrypts the reasoning its models use to arrive at answers, which users cannot access. The operators copied this encrypted reasoning from one conversation and then asked another model in a different conversation to decrypt and transcribe it. However, they did not breach the encryption, access any databases, or retrieve stored user conversations.
In response, OpenAI has banned the involved accounts, strengthened sign-up checks, and implemented protections for hidden reasoning. They also shared their findings with other AI labs through the Frontier Model Forum and relevant government channels.
"Our concern is about violation of our terms of service, not open models or legitimate distillation," Caroline Zier, leading strategic national security policy initiatives at OpenAI, told Bloomberg.
Accusations and Counterarguments
This is the first time OpenAI has directly accused Moonshot, as reported by Bloomberg. Anthropic had made similar claims against the company earlier this month. US agencies have also accused Chinese firms of similar actions, as noted in an advisory that named six firms and the US models targeted. China, however, rejected the advisory as baseless.
Not all agree with these claims. David Sacks, Trump’s former AI czar, has questioned the reports, suggesting they are a ploy to ban rival open models in the US.
Moonshot also faces scrutiny at home, with Chinese regulators investigating DeepSeek and Moonshot, as reported by The Information this month.