This paper discusses the risks associated with large language model (LLM) agents in scientific domains, emphasizing the need for safeguarding over autonomy. While LLM agents show promise in autonomous scientific discovery, they introduce novel vulnerabilities that require careful consideration. The authors propose a triadic framework involving human regulation, agent alignment, and environmental feedback to mitigate these risks. They highlight the potential dangers of misuse, including the creation of hazardous substances, biological risks, and environmental impacts. The paper also identifies vulnerabilities in LLMs, planning, action, external tools, and memory modules, which can lead to harmful outcomes. Current efforts in agent safety include alignment methods, red-teaming, and benchmarking. However, challenges remain in developing specialized models, ensuring domain-specific knowledge, and evaluating safety effectively. The authors advocate for comprehensive safety frameworks, ethical guidelines, and robust regulations to ensure the responsible development and use of LLM agents in scientific applications. The paper underscores the importance of balancing autonomy with safety to prevent misuse and unintended consequences.This paper discusses the risks associated with large language model (LLM) agents in scientific domains, emphasizing the need for safeguarding over autonomy. While LLM agents show promise in autonomous scientific discovery, they introduce novel vulnerabilities that require careful consideration. The authors propose a triadic framework involving human regulation, agent alignment, and environmental feedback to mitigate these risks. They highlight the potential dangers of misuse, including the creation of hazardous substances, biological risks, and environmental impacts. The paper also identifies vulnerabilities in LLMs, planning, action, external tools, and memory modules, which can lead to harmful outcomes. Current efforts in agent safety include alignment methods, red-teaming, and benchmarking. However, challenges remain in developing specialized models, ensuring domain-specific knowledge, and evaluating safety effectively. The authors advocate for comprehensive safety frameworks, ethical guidelines, and robust regulations to ensure the responsible development and use of LLM agents in scientific applications. The paper underscores the importance of balancing autonomy with safety to prevent misuse and unintended consequences.