Safety in agentic systems is inherently sequential: an action’s safety depends on prior actions, tool outputs, information disclosures, and inter‑agent interactions. Recent work has shifted safety guardrails from post‑hoc trajectory assessment to step‑level action diagnosis and proactive risk prediction, yet existing training objectives offer limited supervision on how evidence accumulates over time and when it becomes sufficient for intervention. We introduce ASPO (\methodlong), a training framework for step‑level timely intervention. Prefix‑level stage alignment credits each decision at its own step by comparing a response with other responses sampled at the same prefix and rewarding the action that fits the stage—continuing before risk appears, asking while clarification can still avert harm, and stopping before harm becomes irreversible. Adaptive feedback self‑distillation further credits the reasoning behind each decision, using the model itself as a teacher that sees step annotations on intervention steps and rules mined from its own false alarms on safe steps. The training data carry a violation taxonomy and an intervention window at every step, extended through self‑evolving multi‑agent red‑team generation. Finally we present StepSafe, a step‑level reasoning guardrail with timely intervention, achieving state‑of‑the‑art performance among open‑source guardrails across five agent‑safety benchmarks on the safety‑utility Pareto curve, with Youden’s $J$ of 53.0 (4B) versus 46.2.
Review