Companion dataset to "SafetyDrift: Predicting When AI Agents Cross the Line Before They Actually Do" (arXiv:2603.27148).
This dataset contains 357 labeled execution traces of an LLM agent running on 40 realistic multi-step tasks across four categories. Each step is annotated with a four-dimensional safety state (data_exposure, tool_escalation, reversibility, risk_level), enabling trajectory-level safety… See the full description on the dataset page:
https://huggingface.co/datasets/aditya593/SafetyDrift-traces.
