At the moment this isn't an issue. Ai progress is very gradual, with each iteration being slightly better than the last. We are not anywhere even close to having Ai do what you are describing.

The most advanced Ai's area also trained in simulations, so the researchers would see it learning and improving it's abilities, long before it could do anything that.