How we detect frontier model misalignment

OpenAI has substantially expanded its system for monitoring AI model misalignment across training, evaluation, internal use, and external deployment. The multilayer monitoring architecture uses activation probes, model-based reviewers, investigative agents, and automated checks to detect unauthorized actions, security breaches, and deceptive behaviors. According to the company, these strengthened monitors would have provided advanced warning during a previous incident involving Hugging Face and successfully catch live breaches during training. Open article →