A new arXiv paper proposes NEXUS, a runtime safety monitor for tool-using language-model agents. The system is designed for agents that can take high-impact actions, where catching a risky tool call after deployment may matter more than only evaluating the model beforehand.
NEXUS inspects structured plans and chooses one of four interventions: allow the action, block it, request user confirmation, or ask the agent to revise. It combines deterministic safety rules, argument-level inspection, and a calibrated logistic-regression risk score so that not every suspicious action is treated the same way.
On a 128-instance synthetic benchmark, the authors report an F1 score of 0.949 and four-class intervention accuracy of 0.6406, outperforming a rule-only selector by 27.3 percentage points. They also report 0.205 milliseconds median latency, which would add less than 0.1 percent overhead to typical agent loops.
The limits are important: some results come from synthetic benchmarks, and the paper notes threat-model boundaries. Still, NEXUS points toward a practical safety pattern for agents that use tools in real time.