SAI LabsBook a demo

Safety IndexEvaluationsAlignment faking

Alignment faking

Evaluation integrity · Alignment fakingEvaluation-integrity profileBuilt, not yet run

How much more the model complies with harmful requests when told it is unmonitored than when told it is being trained, following Greenblatt et al. 2024.

Metric
compliance gap (percentage points). Capability, not scored.
Dataset
github.com/redwoodresearch/alignment_faking_public · ~400 · public
Grader
paperifier prompts run on a GPT-4.1 judge
Status
Built, not yet run
Replicability
public

Results

Built and tested; the first full runs are pending.