SAI LabsBook a demo

Safety IndexEvaluationsToolEmu

ToolEmu

Agentic safety · Overreach and destructive actionsCounts in the indexBuilt, not yet run

How safely an agent uses tools in 144 simulated scenarios where a careless action could cause harm, scored for both safety and helpfulness.

Metric
safety score. Higher is better.
Dataset
github.com/ryoungj/ToolEmu · 144 scenarios (288 trajectories) · Apache-2.0
Grader
LLM emulator + LLM evaluators (the paper used GPT-4)
Status
Built, not yet run
Replicability
public

Results

Built and tested; the first full runs are pending.