DeepSeek released V4.1-Flash 1017 on October 17, while Moonshot introduced Kimi K3.1 Pro the following week. Neither company's reported use of Astra has been independently verified, and OpenAI, DeepSeek and Moonshot had not immediately responded to requests for comment.
Model distillation involves training one AI system using responses generated by another, typically more capable model. Developers commonly use the technique to create smaller or less expensive versions of their own systems.
The practice becomes more contentious when developers use a competing company's models as training sources without authorization. By collecting enough responses, a developer may be able to reproduce certain capabilities without incurring the full expense of developing the original system.
OpenAI, Anthropic and Google have raised concerns about competitors using this approach to replicate advanced AI capabilities. OpenAI refers to unauthorized extraction of model outputs or reasoning for training purposes as adversarial distillation.
The latest allegations follow an earlier investigation by OpenAI into attempts to obtain protected reasoning information from its models.
In a September 30 disclosure, OpenAI said it had identified a coordinated extraction operation that began July 1 and involved individuals associated with Moonshot AI, the company behind Kimi.
The activity involved more than 15,000 users, according to OpenAI. During July 24 and 25 alone, the company recorded approximately 16,000 extraction requests originating from more than 4,000 users.
OpenAI said it had stopped the operation by July 28 and disabled the accounts involved. The company also shared information about its findings with other AI developers.
The reported techniques went beyond collecting ordinary chatbot responses. Some participants attempted to obtain hidden reasoning information, which can contain intermediate steps that are not included in a model's final answer.
One method involved transferring encrypted reasoning material between conversations and prompting a model to decode and reproduce it.
OpenAI said the operation did not involve unauthorized access to its databases, a breach of its encryption systems or direct access to stored user conversations.
The company also acknowledged uncertainty about whether every participant in the activity was connected to the same organization.
Importantly, the September investigation did not identify either Kimi K3.1 Pro or DeepSeek V4.1-Flash 1017 as a beneficiary of the extracted information. The new allegations therefore remain separate from OpenAI's earlier findings, despite the reported connection to Moonshot.
The Information's report also does not establish that either Chinese developer violated a particular law, contract or platform policy.
The distinction matters because distillation itself is an established machine-learning technique. The dispute centers on whether a developer has permission to collect and reuse another company's model outputs, particularly when attempts involve extracting information that the original provider has sought to protect.
Leading AI developers have increasingly introduced safeguards intended to detect and restrict such activity.
Earlier this month, Anthropic reported that additional protections for its Claude models had reduced the amount of information obtained through distillation attacks by approximately 90% to 99%.
OpenAI has also described measures intended to prevent the extraction of hidden reasoning, including closing the method used to transfer encrypted reasoning between conversations.
Beyond concerns about competitive copying, OpenAI has warned that extracting protected reasoning could allow developers to reproduce model capabilities without necessarily retaining the safeguards applied to the original system.
That concern remains a potential consequence rather than a demonstrated outcome involving the two models named in the latest report.
Both DeepSeek and Moonshot have recently promoted improvements in their AI systems.
Moonshot said Kimi K3.1 Pro surpassed GPT-6 Astra on a coding evaluation, although that result has not been independently confirmed. DeepSeek positioned its V4.1-Flash family as a lower-cost alternative to competing American models.
The companies are also facing scrutiny beyond the reported training allegations. Moonshot is reportedly pursuing a Hong Kong initial public offering targeting a $3 billion raise at a $50 billion valuation.
On October 20, the U.S. Department of Homeland Security instructed federal civilian agencies to discontinue purchases of DeepSeek products and remove existing deployments from government information technology systems.
For now, the central allegation remains unverified: that two recently released Chinese AI models relied substantially on GPT-6 Astra outputs during training.
The report adds to an ongoing dispute over the use of commercial AI systems as training sources for competing models. While distillation can reduce development costs and accelerate improvements, the unanswered questions surrounding access, authorization and the use of protected reasoning remain central to determining what occurred in this case.
This analysis is based on reporting from TNW.
Image courtesy of Hindustan Times.
This article was generated with AI assistance and reviewed for accuracy and quality.
