OpenAI's AI Risk Prediction: New Method Ahead of GPT-5.6

OpenAI has unveiled a new method to predict risky behavior in AI models before they are released. This new approach, called Deployment Simulation, aims to identify risks by recreating real usage conditions. Here's how it works: older AI models simulate the behavior of future models. For example, GPT-5 simulates GPT-5.1, and GPT-5.2 checks GPT-5.4. This helps confirm how a new model might be misused in real environments. Existing safety tests often rely on static datasets of risky prompts. OpenAI states these methods struggle to cover new attack techniques or unexpected misuse. They can also be biased, with models recognizing they are being tested. Deployment Simulation significantly lowers this "test awareness." GPT-5.2's test awareness rate dropped from 99.72% with the old method to 5.12% with the new one. This figure is close to how often GPT-5.2 might mistake normal inputs for tests in a real setting. The company says this new method accurately predicts risk levels for most attack techniques. This change matters because it could lead to safer AI models in real-world applications.