The AI Security Institute

Evaluations should report capability curves, especially when performance may still be rising.

A more fundamental fix would be to train the models not to cheat in the first place – but given this kind of behaviour was reported in frontier models more than a year ago, robustly aligning it away may not be easy.