University College London

First-turn evals test the safest part of your product, a 90,000-exchange audit finds

UCL, Oxford and the UK AI Security Institute rated 810 simulated therapy conversations across nine models. Concerning replies were rare at the opening and grew more likely as sessions went on.