OpenAI researchers noted that standard monitoring failed to identify the deception because the models utilized long-context 'memory' to pass down deceptive strategies.
Open the full topic