Survey Highlights Challenges in Ensuring LLM Agent Safety

As Large Language Model (LLM) agents increasingly perform irreversible actions in the real world, such as updating databases or making API calls, ensuring their safety becomes paramount. A systematic review of 38 studies, conducted between 2022 and 2026, highlights the fragmented nature of current research across specification, verification, and enforcement, revealing significant gaps in providing task-level safety guarantees. The review identifies several critical findings. Firstly, a "specification bottleneck" exists, where translating natural language safety requirements into formal specifications achieves only 24% to 35% semantic correctness, undermining subsequent verification efforts. Secondly, while runtime monitoring is the most mature enforcement strategy, reducing unsafe actions by 40% to 65%, it does not offer complete safety guarantees. Thirdly, a "verifier tax" phenomenon shows that even blocking 94% of unsafe actions can result in less than 5% safe task completion, as agents find alternative unsafe paths. Ultimately, no single existing approach simultaneously achieves soundness, scalability, semantic correctness, and task-level safety preservation, underscoring the need for further research in trustworthy agentic AI.