Skip to main content
AI RADARModels
177•10/01/2026, 15:02•2 min read

OpenAI Uncovers Mass Distillation Campaign Against Its Models

OpenAI Uncovers Mass Distillation Campaign Against Its Models
Executive TL;DR30s read

OpenAI says it detected a coordinated attempt to extract hidden reasoning from its models, with more than 16,000 requests from over 4,000 accounts at peak. The company has since restricted transfer of encrypted reasoning blocks and strengthened account checks.

OpenAI Uncovers Mass Distillation Campaign Against Its Models

OpenAI said it discovered a coordinated campaign aimed at extracting hidden reasoning from its models. At peak, more than 16,000 such requests passed through the system over two days from over 4,000 accounts.

The company stressed that encryption was not broken: attackers simply moved encrypted chain-of-thought (CoT) blocks from one dialogue to another and asked the model to reproduce them as plain text.

Anthropic previously alleged that Moonshot replaced Kimi's answers with Claude's responses. OpenAI has now published its own investigation. It does not claim Moonshot was behind the entire campaign, but links the main activity cluster to Kimi developers and describes the effort as an attempt to distill its models.

After the investigation, OpenAI closed the ability to transfer encrypted reasoning blocks between users and workspaces, strengthened account verification, and added specific protections against outputting hidden reasoning. The company also shared findings with other AI labs and authorities.

More details are available in the OpenAI report.

Why it matters

Analysis

The case shows how model architecture can become a leakage vector for training rival systems, and sets a precedent for AI labs responding to distillation abuse.

Discuss in community

Share your questions and insights with developers

Join Telegram
+5 Points
← Back to news feed

Related in Ecosystem

Practical guides, core terminology, and agent skills