StepSafe: Learning Timely Intervention with Context-Aware Safety Reasoning
StepSafe is a step-level safety guardrail for agentic systems that learns when to continue, ask for clarification, or stop by modeling how risk evidence accumulates across a trajectory rather than judging only the final outcome. Its main contribution is a training setup with prefix-level stage alignment, adaptive self-distilled reasoning feedback, annotated intervention windows, and self-evolving multi-agent red-teaming, which is a nice fit for sequential safety decisions, though the claims seem heavily benchmark-driven and may depend on the quality of mined rules, annotations, and red-team coverage.
Abstract
Safety in agentic systems is inherently sequential: whether an action is safe depends on prior actions, tool outputs, information disclosures, and interactions among agents. Recent work has shifted agent safety guardrails from post-hoc trajectory-level assessment toward step-level action diagnosis and proactive risk prediction, yet existing training objectives provide limited supervision for how step evidence accumulates over time and when it becomes sufficient to justify intervention. We propose \methodlong~(ASPO): a training framework for step-level timely intervention. Pre-fix level stage alignment allows crediting each decision at its own step: it compares each response with the other responses sampled at the same prefix and rewards the action that fits the stage, continuing before the risk is visible, asking while clarification can still avert harm, and stopping before harm is irreversible. Adaptive feedback self-distillation further credits the reasoning behind each decision, with the model itself as a teacher that sees the step annotations on intervention steps and rules mined from its own false alarms on safe steps. Its training data carry a violation taxonomy and an intervention window at every step and are extended through self-evolving multi-agent red-teaming. Finally we present \textsc{StepSafe}, a step-level reasoning guardrail with timely intervention, that achieves the state-of-the-art peformance among open-source guardrails across five agent-safety benchmarks on safety-utility pareto curve with Youden's $J$ of 53.0 (4B) against 46.2.