OpenAI’s new safety tools are designed to make AI models harder to jailbreak. Instead, they may give users a false sense of security