OpenAI blocked an organised attempt to extract large amounts of non-public reasoning information from its AI models. The company judged that the core group behind the activity was linked to people connected to Chinese AI startup Moonshot AI.

On Oct. 1, CNBC reported that the activity disclosed by OpenAI began on July 1 on a small scale. On July 24-25, OpenAI saw 16,000 requests showing similar extraction patterns from more than 4,000 users. Further investigation confirmed related activity in a group of more than 15,000 users, and OpenAI blocked it by July 28. The 16,000 figure reflects attempts, not successful extractions of reasoning.

OpenAI called it "adversarial distillation". It refers to obtaining another AI's outputs or reasoning in bulk without permission and using it to train a separate model or improve performance. Model distillation itself is widely used, but OpenAI said it becomes an issue when protected features of a competitor are copied without authorisation.

The incident differed from traditional hacking that breaks into servers or databases. The operators could not decrypt OpenAI's encryption or directly access stored user conversations. Instead, they tried to exploit model interactions so that non-public reasoning would be reproduced in a form visible to users. OpenAI blocked related pathways, restricted suspicious accounts and strengthened its sign-up and monitoring systems.

There remains room for interpretation in assessing who was behind the activity. OpenAI said it could not confirm whether all operators came from a single organisation, but added that the core activity group was attributable to people related to Moonshot AI, which developed the AI service Kimi. It said it shared its findings with other AI companies through the Frontier Model Forum and government information-sharing channels.

Anthropic has previously claimed that Chinese AI firms, including Moonshot AI, sought to use Claude's reasoning capabilities to improve their own models. Anthropic said it observed more than 23 million exchanges related to Moonshot AI from May to July this year. Moonshot AI did not immediately respond to inquiries about OpenAI's announcement.

As AI models' reasoning capabilities emerge as a key competitive asset, competition among AI companies over security is expected to intensify further, including attempts to replicate them and defensive technologies.