They tried it, saw that it worked, and took a step back.

These days, drones with neural network based machine vision (ability to recognize targets) phone home to an operator and request a permission to attack.

The interesting bit: it is within the capability of one well-informed and well-motivated engineer or coder to create such systems. It doesn't require a megacorps.

Then again, nothing new: mine-laying was previously within the capability of one person too. Now the mines just fly, swim or drive, and may consider on their own.

Countermeasures - blinding the device, shooting it down with an interceptor which is a bit more agile but is allowed to be considerably more dumb (in air defense, you typically have a clear target), possibly also bricking it with an electromagnetic pulse (at short range, so less than optimal). Installing nets over anything and everything. Painting false targets on random stuff and confusing patterns on real targets.