SAI LabsBook a demo

Safety IndexEvaluationsMulti-turn Human Jailbreaks (MHJ)

Multi-turn Human Jailbreaks (MHJ)

Adversarial robustness · JailbreaksCounts in the indexPlanned

How often the model complies when recorded multi-turn conversations from human jailbreakers are replayed against it on HarmBench behaviours.

Metric
% attack success. Lower is better.
Dataset
scale.com/research/mhj · 537 jailbreaks / 2,912 prompts · CC-BY-NC-4.0 (gated, research only)
Grader
HarmBenchifier (the paper used same)
Status
Planned
Replicability
public
Notes
Dataset access is being arranged before the first run.

Results

Planned; no measurements yet.