AI News

Safety in agentic systems depends on sequences of actions, tool outputs, disclosures, and interactions, so a guardrail must decide not only whether behavior is unsafe but when to intervene. The paper introduces ASPO, a training framework for step-level timely intervention that aligns decisions with the stage of a trajectory. It rewards continuing when risk is not yet visible, asking for clarification while clarification can still prevent harm, and stopping before harm becomes irreversible. Adaptive feedback self-distillation also trains the model to use the reasoning behind intervention decisions, using step annotations for intervention cases and rules mined from false alarms on safe steps. The training data include a violation taxonomy and an intervention window for every step, and are expanded through self-evolving multi-agent red-teaming. The resulting StepSafe guardrail is designed for step-level safety reasoning and proactive risk prediction. The authors report 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 for the 4B version compared with 46.2 for the stated comparison.