So pessimistic.
We don't need to know how brains work to have useful interactions with them, even pulling out pictures and words from neural activity.
That sensor breakthrough has already happened- and no dangerous intermediate step was necessary. opnwatr can make 8 million 100micron wide scans 120 times per second with what is basically a slightly modified version of cheap commodity/consumer hardware (using IR LEDs).[1] fMRI is slower and lower resolution and can still be used to extract images from brains in real-time.