"We made a mistake in not being more transparent": OpenAI secretly accessed benchmark data, raising questions about the AI model's supposedly "high scores" — after Sam Altman touted it as "very good"