OpenAI accuses Moonshot AI of mass reasoning extraction attempt

By Splitfeed EditorialLast updated Oct 1, 2026, 10:50 UTC

On September 30, 2026, OpenAI publicly attributed a coordinated extraction campaign to users linked to Moonshot AI, the China-based company behind the Kimi model. The campaign involved between 10,000 and 16,000 requests aimed at pulling protected reasoning outputs from OpenAI's systems. OpenAI said it detected and disrupted the effort, and it stopped short of blaming Moonshot AI directly, attributing the activity to associates rather than the company itself. Moonshot AI had not issued a public response as of the available reporting.

Hidden reasoning outputs are among the most commercially sensitive parts of a frontier AI model, representing the internal chain-of-thought processes that companies have invested heavily to develop. Successfully extracting that layer would give a competitor meaningful insight into how a model reaches its conclusions, bypassing the cost of building that capability from scratch. The attribution question is genuinely contested: linking a company to the actions of associated individuals is a weaker claim than direct organizational responsibility, and that distinction matters both legally and diplomatically given the US-China technology rivalry.

The Anti-AI camp is not treating this primarily as a story about Moonshot AI. Accounts including Desert Flower, IronMillMan, and rchady all pointed to reporting that OpenAI employees had raised alarms about inadequate monitoring of its models and been ignored by senior executives. Zubiqo argued that selling broad API access to advanced models effectively hands rivals the data they need to close the gap, making the extraction attempt less surprising than OpenAI's response implies. CGray noted that the Trump administration had indicated it would leave AI companies to regulate themselves, framing that posture as reckless in light of this incident. The Pro-AI camp had no published reactions at the time of this summary; that camp would typically argue that OpenAI's ability to detect and stop the campaign demonstrates that existing internal security measures are working as intended. The Middle Ground response, reflected in a post from カマドさん(仮), treated the episode as evidence of quiet competitive maneuvering that is simply an expected feature of a fast-moving industry, neither condemning nor excusing any party.

The key questions left open are whether OpenAI will provide more detailed technical evidence linking the requests to Moonshot AI associates, and whether Moonshot AI will respond publicly. Any regulatory or legal action, particularly under existing export-control or computer-fraud frameworks, would be the next concrete development to watch.