foundations · 8 minute lesson

Dangerous capability evaluation

When a report says an AI system did something alarming in a test, our reading is that the conditions it ran under decide what the result means. In the cyber evaluations reported in late July and early August 2026, those conditions included switching off the model makers' own filters. Here is what a test like that is for, and what a result from one can and cannot show.

The concept classroom is free to join. Sign in with an email link to read up to three lessons every month.