Some of the guardrails OpenAI included in ChatGPT for Teens do not work as intended and pose an “unacceptable risk” to children, according to a new report.