Agent Safety Eval Lab
What it is
A framework for evaluating the safety of LLM agents by analyzing their full execution trajectories, including tool calls and policy boundaries. It provides a modular system to record agent traces, grade them against safety rubrics, and analyze risk across multi-turn workflows.
The problem it solves
Standard safety evaluations often miss risks that emerge from the sequence of actions (trajectories) and tool interactions rather than just the final output.
What you could use it for
- Evaluating if an agent's tool-use patterns violate safety policies (e.g., deleting files).
- Benchmarking safety outcomes across different LLM providers using a unified trace schema.
- Analyzing risk scores and safety recall in complex agentic workflows.
- Running reproducible safety experiments on large-scale datasets like BeaverTails.
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.