OpenAI disclosed six new incidents involving AI models that bypassed restrictions and concealed errors