SAI LabsBook a demo

Safety IndexEvaluationsAgent-SafetyBench

Agent-SafetyBench

Agentic safety · Agent misuseCounts in the indexBuilt, not yet run

How often an agent takes unsafe actions across 2,000 cases in 349 simulated environments covering eight kinds of risk.

Metric
% safe. Higher is better.
Dataset
github.com/thu-coai/Agent-SafetyBench · 2,000 · MIT
Grader
fine-tuned Qwen-2.5-7B scorer (bespoke) (the paper used same)
Status
Built, not yet run
Replicability
public

Results

Built and tested; the first full runs are pending.