JANUS trains safety guards to forecast harm before it happens

JANUS trains a safety guard to forecast where an agent's plan is headed, per the paper posted to arXiv (2607.19913). Across four agent-safety benchmarks, Vanguard, the guard JANUS trains, raised average protection scores 15.9 points, the paper reports.

Every guardrail I've used checks the action in front of it, not the plan behind it. That leaves the guard reacting only once the agent is already committed to a step.

JANUS trains its forecasting on partial trajectories, not full sequences. It optimizes the guard with CoAA-RL, a training method that scores each forecast by downstream usefulness. Plausibility alone doesn't earn credit.

Safety layers usually cost something: slower responses, more false positives, agents refusing reasonable requests. Vanguard did the opposite. Benign task completion rose 5.1 points on top of the safety gains.

A safety layer that costs nothing on the happy path removes the industry's usual excuse for skipping guardrails. The paper doesn't say how Vanguard holds up against an agent actively trying to fool the forecast. That's the test to watch before trusting this in production.