Overview

• OpenAI has introduced the Safety Evaluations Hub, a platform to share safety metrics for its AI models, including tests for harmful content, jailbreaks, and hallucinations.
• The hub will be updated regularly with safety data and will include evaluations for all major model updates moving forward.
• OpenAI plans to expand the hub’s scope over time, potentially adding more safety tests as evaluation methods advance.
• The initiative comes in response to criticism from ethicists about rushed safety testing and follows the rollback of GPT-4o due to problematic outputs.
• An opt-in alpha testing phase will allow select users to provide feedback on new models before their official release, aiming to prevent future issues.