Home ยป AI Security Assessment Framework Introduced by OpenAI for Pre-deployment Evaluation of Artificial Intelligence

AI Security Assessment Framework Introduced by OpenAI for Pre-deployment Evaluation of Artificial Intelligence

OpenAI has recently announced the framework for preparedness in order to assess the safety of AI models. This framework aims to provide a clearer and more structured process for evaluating whether AI models are safe or not.

OpenAI already has a dedicated team responsible for AI safety. This team is responsible for overseeing and evaluating different AI models, categorizing them into three groups: well-aligned models that are considered safe, models in the preparedness stage that need to be assessed according to the framework before being deemed safe, and high-ability models that may pose potential risks in the future.

The Preparedness Framework utilizes scoring cards to assess and rate models on a scale of low, medium, high, and severe. The four dimensions of safety that are considered include cybersecurity, CBRN (chemical, biological, radiological, and nuclear) safety, AI bias, and model autonomy. In order to be deemed safe, the overall score must be either low or medium. If the score falls in the high range, adjustments must be made before the AI model can be considered safe for deployment.

Once a model has been scored within the framework, the AI safety team will pass it on to the core team for further evaluation. The final decision of whether or not to deploy the AI model lies with the management, with the ability to halt the process if they perceive any potential risks.

For more detailed information about the Preparedness Framework, please refer to the link provided by OpenAI.

TLDR: OpenAI has introduced a framework for assessing the safety of AI models, categorizing them into well-aligned, preparedness, and high-ability groups. The Preparedness Framework utilizes scoring cards to evaluate models based on cybersecurity, CBRN safety, AI bias, and model autonomy. Models must receive a low or medium score to be considered safe. The final decision to deploy AI models lies with the management, with the ability to halt the process if they perceive any risks.

More Reading

Post navigation

Leave a Comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Nearly 6 out of 10 victims of sexual crimes in South Korea are minors targeted by Deepfake technology.

BlackBerry Abolishes IoT IP Business Blueprint Concurrently Appointing New CEO

The United States Department of State Accuses SolarWinds’ CISO of Negligence; Issues Comprehensive Report on Company’s Impacted Operations Due to Hacking Incident.