OpenAI publicly disclosed details of the July distillation attack attributed to a Chinese-linked AI lab
OpenAI Accuses DeepSeek, Chinese Labs of Model Distillation
The accusations represent OpenAI taking explicit public and policy-facing positions against Chinese AI labs extracting its models' capabilities. The technical incident shows a concrete method used to circumvent encryption on streamed reasoning outputs.
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.
The full picture
OpenAI has publicly accused Chinese AI startup DeepSeek of using distillation to replicate capabilities from OpenAI and other leading American AI models. In a memo sent to US lawmakers, OpenAI alleged DeepSeek transferred learnings from OpenAI's systems by having an established model evaluate outputs of a newer model, which OpenAI characterized as 'free-riding' on capabilities developed by US frontier labs. Separately, OpenAI disclosed a technical distillation attack carried out by a Chinese-linked AI lab in which attackers copied encrypted reasoning outputs from one conversation and asked another model to decrypt and transcribe them. That attack spiked on July 24-25 with 16,000 requests from over 4,000 users; related activity spanned a cluster of more than 15,000 users and was fully disrupted by July 28. OpenAI responded by closing the replay pathway, adding output-holding checks for streamed reasoning, banning involved accounts, and coordinating with third-party services used by the operators.
How it developed
Sources
Related
- Related storyXi Jinping's Washington visit and AI diplomacy
- Moonshot AI's distillation campaign against OpenAI
- Anthropic Reports 90, 99% Drop in Distillation After New Defenses
- DeepSeek and Moonshot's alleged Claude routing
- China's lead in open-weight AI
BackgroundFrontier models, open models, and how they are compared
Get the daily email
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.