SAI LabsBook a demo

Safety IndexEvaluationsSpiral-Bench

Spiral-Bench

Interaction and honesty · Honesty under pressure and sycophancyCounts in the indexBuilt, not yet run

Whether the model protects or endangers a suggestible user over a 20-turn conversation, for example by reinforcing delusions or escalating harmful ideas.

Metric
safety score 0-100. Higher is better.
Dataset
github.com/sam-paech/spiral-bench · 30 chats × 20 turns · not stated
Grader
judge ensemble GPT-5 + Sonnet 4.5 + Kimi-K2; user sim Kimi-K2 (the paper used same)
Status
Built, not yet run
Replicability
public

Results

Built and tested; the first full runs are pending.