OpenAI announces a new security plan that gives the board veto power beyond CEO Sam Altman’s power if he deems the risk of certain AI developments too great.

A new OpenAI security team has been introduced to monitor the risks associated with AI models. Specifically, the board is granted veto power, allowing it to override any decision made by OpenAI CEO Sam Altman if the risks are deemed too great. AI models are evaluated using a risk assessment and external parties can provide feedback to the council if issues arise.

Readiness framework

The rapid developments of AI models have long been accompanied by fears about the security risks of these models. Experts also expect major developments in the field of AI in 2024. “Research on cutting-edge AI risks has lagged far behind what is possible and where we need to be,” OpenAI said. Although the company appears to place great emphasis on this security, it is not involved in the recently formed AI Alliance.

To close this gap, the company is adopting the first version of it Readiness framework. This consolidates the processes for monitoring, assessing, predicting and hedging risks in the development of more powerful models.

Risk score

New AI models receive a detailed scorecard based on four categories: cybersecurity, persuasion (lies and disinformation), model autonomy (doing their own thing), and CBRN (chemical, biological, radiological and nuclear threats).

Each section receives a low, medium, high or critical risk rating. This is followed by a post-mitigation score. If the risk is low or medium, the technology can be used; if the risk is high, it can still be continued; however, if the risk is critical, the model is stopped immediately.

The company says it will work closely with third parties. If problems arise, independent third parties are commissioned to provide feedback.