Researchers have conducted a survey to identify safe methods for Large Language Model (LLM) agents. They analyzed 38 studies and found that current approaches to specifying, verifying, and enforcing safety guarantees are fragmented and often ineffective. The main challenges include translating natural language into formal specifications, which is only around 25% accurate, and runtime monitoring, which can reduce unsafe actions but not provide complete safety guarantees. A new taxonomy and research agenda have been proposed to address these issues. --- Why it matters: This study matters because it highlights the need for more effective methods to ensure the safety of LLM agents, which are increasingly performing critical tasks in real-world settings. Researchers and engineers working on AI systems will be interested in understanding the limitations of current approaches and exploring new solutions. Source: https://arxiv.org/abs/2608.14590