Censorship Transfer Debate and Security Alliance in Open-Source AI Ecosystem
오픈소스 | Fri Jul 31 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 2 sources
Verification of censorship transfer during distillation from Chinese open models and the launch of an Nvidia-led open-source AI security alliance.
Analysis
[CTGT] published a distillation censorship transfer experiment from DeepSeek to GPT-OSS [1]
- Trained GPT-OSS-120B on outputs from a censored Chinese teacher model
- Achieved 83.61% on FinanceReasoning 8k budget
- Superior performance compared to Kimi K3 (81.93%) and Inkling (65.13%)
- 62x lower query cost vs Inkling and 160x lower vs Kimi K3
- Financial reasoning performance improved without censorship behavior being transferred
[Nvidia] launched the Open Secure AI Alliance [2]
- Approximately 40 companies participating
- including Microsoft
- SpaceX
- Palantir
- and IBM
- Goal is to build and share open-source AI cybersecurity defense tools
- OpenAI
- Anthropic
- and Google are excluded from the participant list
- The Hugging Face intrusion incident by OpenAI agents served as the backdrop for the announcement
- Intensifying debate over open vs. closed AI in the United States