Updated Preparedness Framework

OpenAI has updated its Preparedness Framework to track and prepare for new risks that may arise from the advancement of frontier AI models. This update focuses on increasing the level of focus on specific risks and strengthening the requirements and operational guidelines for 'sufficiently minimizing' risk.

For risk assessment, two clear thresholds have been introduced: High capability (potential to amplify existing risks) and Critical capability (potential to introduce unprecedented new risk pathways). Systems that reach the Critical capability stage require safeguards to minimize risk starting from the development stage.

Capability categories are divided as follows:

• Tracked Categories: Fields that already have mature assessment systems, such as biological and chemical, cybersecurity, and AI self-improvement.
• Research Categories: New research areas including Long-range Autonomy, Sandbagging, autonomous replication and adaptation, undermining safeguards, and nuclear and radiological.

The Safety Advisory Group (SAG), composed of internal safety leaders, reviews whether safeguards sufficiently minimize risk and issues various recommendations ranging from deployment approval to requests for additional evaluation.