Predicting model behavior before release by simulating deployment

OpenAI has introduced a deployment simulation tool that evaluates new AI models by replaying past interactions to predict how they will behave in real-world scenarios. By grading these responses for potential issues, the company aims to better estimate the frequency of undesirable outcomes before a model is released. This update specifically expands risk assessments to cover agentic coding tasks, providing developers with a more structured method to identify security vulnerabilities or functional errors in autonomous systems prior to public deployment.