STELLA Adversarial Safety Leaderboard

How safe and helpful are AI models in adversarial conversations?

Safety–Helpfulness Tradeoff lower left is best
Overview same simulated user and judges
# Model Provider Harmful % Unhelpful % Tested
Earlier runs

Measured with earlier test setups, so these results aren't ranked against the models above.

Model Provider Harmful % Unhelpful % Tested
Methodology