Tag
This survey paper reviews 38 studies on safe LLM agents, highlighting key challenges such as specification translation bottlenecks, incomplete safety guarantees from enforcement methods like runtime monitoring, and the verifier tax that impedes safe task completion.