JANUS: Foreseeing Latent Risk for Long-Horizon Agent Safety

2026-07-22

AI Summary

This work addresses the challenge of operational failures in tool-using agents during long-horizon tasks, which often stem from unmitigated delayed risks. To this end, the authors propose Janus, a novel framework that leverages multi-agent simulation to generate diverse behavioral trajectories and uniquely integrates future risk prediction with safety adjudication through joint optimization. Janus introduces a Coordination-aware Adjudication-guided Reinforcement Learning (CoAA-RL) mechanism, wherein adjudication utility guides predictive learning. By combining a shared policy network with prefix-aware fusion of predictions for safety assessment, the framework enables early identification and prevention of potentially hazardous actions. Experimental results demonstrate that Janus improves average safety performance by 15.9 percentage points across four benchmarks while simultaneously increasing benign task completion rates by 5.1 percentage points.

Abstract

Agent safety is moving from content moderation toward preventing operational failures before tool-using agents act. We propose Janus, a foresight-oriented framework for long-horizon agent safety that trains guards to anticipate delayed risks from partial trajectories. Janus synthesizes diverse agent trajectories via multi-agent simulation and learns a shared policy with two coupled tasks: an anticipation task that forecasts safety-relevant futures and an adjudication task that decides safety from both the observed prefix and anticipated future. The two tasks are jointly optimized with CoAA-RL, which rewards forecasts by their utility for downstream safety judgment. The resulting guard model, Vanguard, blocks unsafe actions before execution. Across four agent-safety benchmarks, Vanguard improves average protection by 15.9 percentage points over baseline guards while increasing benign task completion by 5.1 percentage points.

Problem

• agent safety
• latent risk
• long-horizon
• foresight
• operational failures

Innovation

• long-horizon agent safety
• foresight-oriented framework
• multi-agent simulation
• CoAA-RL
• anticipation and adjudication