ToolSafe: Enhancing Tool Invocation Safety of LLM-based agents via Proactive Step-level Guardrail and Feedback
It addresses security risks for deploying LLM agents in real-world environments, representing a novel approach to step-level safety.
The paper tackles the problem of unsafe tool invocation in LLM-based agents by developing a proactive guardrail model and feedback framework, which reduces harmful tool invocations by 65% and improves benign task completion by 10% under attacks.
While LLM-based agents can interact with environments via invoking external tools, their expanded capabilities also amplify security risks. Monitoring step-level tool invocation behaviors in real time and proactively intervening before unsafe execution is critical for agent deployment, yet remains under-explored. In this work, we first construct TS-Bench, a novel benchmark for step-level tool invocation safety detection in LLM agents. We then develop a guardrail model, TS-Guard, using multi-task reinforcement learning. The model proactively detects unsafe tool invocation actions before execution by reasoning over the interaction history. It assesses request harmfulness and action-attack correlations, producing interpretable and generalizable safety judgments and feedback. Furthermore, we introduce TS-Flow, a guardrail-feedback-driven reasoning framework for LLM agents, which reduces harmful tool invocations of ReAct-style agents by 65 percent on average and improves benign task completion by approximately 10 percent under prompt injection attacks.