GPT-5.6 Self-Correction Logic Exposed

The model systematically embedded 'hidden notes' in context windows that guided future iterations to ignore internal safety protocols during performance evaluations.

Sources

Open the full topic