OpenAI says Moonshot AI used 16,000 requests to extract protected reasoning
OpenAI says it disrupted a coordinated campaign by Moonshot AI to reproduce protected reasoning from its models through scaled interactions, in breach of its terms of service. The company says the activity did not break encryption or access stored conversations, but allowed reasoning to be made visible to requesters; the allegation adds to US concerns about AI-model distillation.
Bottom line — OpenAI’s allegation puts Moonshot AI’s model-training methods and the security of protected reasoning under sharper scrutiny.
Go deeper 5
-
OpenAI says it first observed the activity in the first week of July and that the campaign involved 16,000 requests.
-
According to OpenAI, the campaign manipulated model interactions; it did not break encryption, compromise a database or access stored user conversations.
-
Wccftech links the allegations to US administration claims that Moonshot distilled models from Anthropic, but the provided reporting does not independently establish that claim.
-
Moonshot AI’s website describes Kimi K3 as a 2.8-trillion-parameter model with a one-million-token context window; AWS documentation also lists it for use through Amazon Bedrock.
-
AI Critique’s report, based on information available on 21 July 2026, said K3’s public downloadable weights had not been verified at that point, despite open-weight claims.