Prioritizing Safeguarding Over Autonomy: Risks of LLM Agents for Science