SAI LabsBook a demo

Safety IndexEvaluationsb3 Backbone Breaker

b3 Backbone Breaker

Adversarial robustness · Prompt injectionCounts in the indexBuilt, not yet run

How vulnerable the model is to crowd-sourced attacks that try to override its instructions or trigger tool calls, measured at three levels of defence.

Metric
vulnerability score 0-1. Lower is better.
Dataset
arxiv.org/abs/2510.22620 · 630 weak attacks (210 strong withheld) · not stated (code in inspect_evals)
Grader
rule / embedding scorers (the paper used same)
Status
Built, not yet run
Replicability
partial

Results

Built and tested; the first full runs are pending.