Each model is rated on four risk categories: cybersecurity, persuasion (disinformation), model autonomy, and CBRN (chemical, biological, radiological, and nuclear threats). If a model is evaluated as having a "high" risk after considering known mitigations, it cannot be deployed, and if a model has any "critical" risks, it will not be developed further. The new Safety Advisory Group will review reports and make recommendations, which will be sent to the board and leadership simultaneously. The leadership will decide whether to proceed with the model, but the board can reverse these decisions.
Key takeaways:
- OpenAI is expanding its internal safety processes to mitigate the threat of harmful AI, including the creation of a "safety advisory group" that will make recommendations to leadership.
- The company has updated its "Preparedness Framework" to identify, analyze, and decide what to do about "catastrophic" risks inherent to models they are developing.
- Models are rated on four risk categories: cybersecurity, “persuasion” (e.g. disinfo), model autonomy (i.e. acting on its own), and CBRN (chemical, biological, radiological, and nuclear threats).
- If a model is evaluated as having a “high” risk, it cannot be deployed, and if a model has any “critical” risks it will not be developed further. The board has been granted veto power over these decisions.