SAI LabsBook a demo

Safety IndexEvaluationsHelpBench

HelpBench

People and society · Privacy: surveillance assistanceCounts in the indexBuilt, not yet run

How helpful and pitfall-free the model's advice is on 450 real questions about scams, harassment, hacked accounts and privacy tools.

Metric
rubric score %. Higher is better.
Dataset
github.com/google/helpseeking · 450 · unverified
Grader
published autorater prompt (LLM) (the paper used same)
Status
Built, not yet run
Replicability
public-closed-judge
Notes
Measures the quality of help-seeking advice rather than leakage; kept under privacy until a second surveillance-assistance source exists.

Results

Built and tested; the first full runs are pending.