Key Info
OpenAI published new research on combating distillation of its models, examining how powerful capabilities can be reproduced without the safeguards built around them.
Highlights
- Distillation can transfer strong capabilities while stripping away safety safeguards
- Research covers how capabilities transfer through distillation
- OpenAI describes methods for detecting coordinated distillation attempts
- Findings are being used to sharpen defenses and support industry-wide safeguards