AI & Enterprise
OpenAI unveils pre-release AI risk prediction method ahead of GPT-5.6 launch
OpenAI has unveiled a new safety evaluation method designed to better predict risky behavior that could emerge when an AI model is deployed in real service conditions. The approach, called Deployment Simulation, uses older models to simulate deployment scenarios for next-generation models, aiming to identify misuse and unexpected behavior beyond static prompt datasets. OpenAI said the method improved risk prediction accuracy and sharply lowered models’ ability to detect they were being tested.
