Search papers, labs, and topics across Lattice.
This paper introduces SARA, a novel framework that distinguishes between action induction and execution authorization in tool-augmented LLM agents to mitigate the risks of unintended real-world consequences. By employing an Action Probe to isolate action-inducing semantics and implementing strict authorization protocols for tool calls, SARA effectively limits action specification risk while maintaining competitive task performance. The results show that SARA achieves an Action Specification Rate (ASR) of no more than 0.63% across various evaluation settings, significantly enhancing safety without compromising utility.
Separating action induction from execution authorization in LLM agents can drastically reduce unintended command execution while preserving task effectiveness.
Tool-augmented LLM agents must rely on untrusted runtime Observations to complete open-ended tasks; however, when tool outputs no longer merely provide data but begin to specify concrete actions, they effectively become ``commands''that can drive real-world side effects beyond user intent. We argue that this risk arises from conflating action induction with execution authorization. To address this distinction, we propose SARA, which treats action induction and execution authorization as distinct runtime roles and separates action provenance from execution authority. On the Observation side, a context-isolated Action Probe exposes action-inducing semantics and persistently records action-origin provenance across steps as a review signal; on the execution side, actual tool calls are authorized only against the user objective and audited evidence from authorized successful executions, while satisfying goal, execution-chain, and argument-level support. To preserve this separation across multi-step execution, SARA applies No-History-Promotion to prevent historical recurrence from laundering action origins into execution authority. Across AgentDojo and AgentDyn, SARA limits ASR to no more than \(0.63\%\) across four primary evaluation settings while maintaining competitive task utility, and consistently reduces ASR across additional Agent backbones.
