Guard models have become popular for constraining LLM‑based agents, mainly by flagging actions that are prohibited. However, detecting prohibited actions alone is insufficient for full safety. We argue that safety also depends on spotting obligations—actions that are required but left undone.
A preliminary analysis on a widely used safety benchmark shows that 56.92% of GLM‑5.3 trajectories contain unfulfilled obligations, while only 30.00% contain prohibited actions. This reveals unfulfilled obligations as a major, previously overlooked safety risk. No existing benchmark evaluates a guard model’s ability to identify such obligations.
To close this gap we introduce ObligationBench, the first benchmark dedicated to obligation identification. It comprises 240 expert‑validated trajectories covering issue resolution, feature development, and terminal operations. Evaluation of 14 representative models shows the best recall and exact‑match rates are only 48.97% and 10.00%, respectively.
To address these limitations we develop ObligationGuard, trained on 40,000 synthetic examples. ObligationGuard achieves 57.52% recall and a 21.67% exact‑match rate, surpassing all evaluated models on both metrics.
We call on the community to incorporate obligation identification into the design and evaluation of future guard models to improve agent safety.
Review