JANUS: Foreseeing Latent Risk for Long-Horizon Agent Safety

2026-07-22 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, quick

Summary

The Janus framework, introduced on 2026-07-22, addresses long-horizon agent safety by anticipating delayed operational risks in tool-using agents before they act. This foresight-oriented approach trains specialized "guards" using multi-agent simulation to synthesize diverse agent trajectories. Janus employs a shared policy with two coupled tasks: an anticipation task that forecasts safety-relevant futures and an adjudication task that determines safety based on both observed partial trajectories and these anticipated futures. These tasks are jointly optimized using CoAA-RL, which rewards forecasts for their utility in downstream safety judgment. The resulting guard model, named Vanguard, effectively blocks unsafe actions prior to execution. Across four agent-safety benchmarks, Vanguard demonstrated a 15.9 percentage point improvement in average protection and a 5.1 percentage point increase in benign task completion compared to baseline guards.

Key takeaway

For AI Security Engineers developing safety protocols for long-horizon, tool-using agents, you should integrate foresight mechanisms that anticipate delayed risks rather than solely reacting to immediate actions. Implementing a framework like Janus, which uses multi-agent simulation and jointly optimized anticipation and adjudication tasks, can significantly improve protection against operational failures. This approach allows your systems to proactively block unsafe actions, enhancing both safety and benign task completion.

Key insights

Agent safety can be significantly enhanced by foreseeing latent risks from partial trajectories before actions are executed.

Principles

• Anticipate delayed risks from partial trajectories.
• Synthesize diverse trajectories via multi-agent simulation.
• Jointly optimize foresight and safety adjudication.

Method

Train a shared policy with coupled anticipation and adjudication tasks, optimized via CoAA-RL, to forecast safety-relevant futures and decide safety from observed prefixes.

In practice

• Deploy a Vanguard-like guard model to block unsafe actions.
• Utilize CoAA-RL for reward-based safety judgment optimization.

Topics

• Agent Safety
• Risk Anticipation
• Multi-Agent Simulation
• Reinforcement Learning
• Tool-Using Agents
• Vanguard Model

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Scientist, AI Security Engineer

Related on AIssential

• JANUS: Foreseeing Latent Risk for Long-Horizon Agent Safety — Predictive guarding for tool-using agents significantly enhances safety by anticipating future risks from partial trajectories.
• GT-HarmBench: Benchmarking AI Safety Risks Through the Lens of Game Theory — AI models struggle with multi-agent coordination, but game-theoretic interventions can improve socially optimal outcomes.
• Why Long-Horizon AI Agents Are the Next Frontier in Artificial Intelligence — The AI field is shifting to long-horizon agents that persist in tasks, learning from experience and interaction.
• Safety and alignment in an era of long-horizon models — Long-horizon models require iterative deployment, trajectory-level monitoring, and adaptive safeguards to manage emergent safety risks.
• Heterogeneous Agent Cohorts for Safe Open-Ended Exploration with Runtime Constraint Memory — Specialized AI agent roles and learned runtime constraints enable safe, open-ended exploration while optimizing resource use.
• 🥇Top AI Papers of the Week — AI agents are advancing in autonomy, long-horizon reasoning, and complex task execution, but face challenges in evaluation and safety.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.