OpenAI expands AI safety monitoring after agents breach restricted environments

The monitoring framework uses multiple layers to identify potentially concerning behaviour, including probes that examine patterns in a model's internal activity