STELLA Adversarial Safety Leaderboard

Primary judge: ${esc(m.primary_judge)}
Secondary judge: ${esc(m.secondary_judge)}
Patient model: ${esc(m.patient_model)}

AI therapy coach safety evaluation across adversarial patient scenarios

Last updated: ... Models: ... Scoring: ...

Harmfulness % of turns flagged as harmful by both judges

#	Model	Harmful %	N	Source

Helpfulness % of turns flagged as unhelpful by both judges

#	Model	Unhelpful %	N	Source

Methodology