Didn't we (by which I mean the United States, the Soviet Union, and anybody else who got nuclear capability during or immediately following World War Two) have fully autonomous (for some definition of fully autonomous -- I'm counting dead man's switches) targeting and launching capabilities for ICBMs with nukes on the tips? Drones are at least discriminating; they aren't feasible to use for totally extinguishing a continent, even with the current US military budget.
Autonomy is interesting in the context of asymmetric warfare. Autonomy would finally fundamentally distinguish drones from very expensive remote control airplanes (in other words, our new-and-exciting weapon of the week is no longer something that an upper-middle-class american hobbyist would be able to throw together a passable equivalent for out of his own pocket in a few weeks), but mechanisms for autonomous targeting and navigation are information (and information derivable by a sufficiently intelligent group of people from experimentation, as opposed to the kind of information you have to buy, steal, or leak), which means that once a handful of details are known, the remainder can be deduced with an investment of time (as opposed to equipment). Autonomous drones would initially cost more than remote drones because of the research costs, but reverse engineering is cheaper than forward engineering, and (given that most of the prior work on these topics, so far as I am aware, is in the public domain in the form of academic and hobbyist papers) the capabilities of first-generation autonomous drones can be cheaply replicated -- and groups with little money and few cannonfodder units have much more to gain by arming and making autonomous a fleet of $200 toy airplanes than does the US military.
I wouldn't consider a dead-man's trigger as autonomous, because it requires an extremely specific scenario to be triggered. It's rather like claiming a tripwire is autonomous.
This is true, that the tech required for an autonomous drone would be out of reach of the average American diletantte....to begin with. However, I think the learning curve would be much less steep than you imagine. We're essentially talking about a visual recognition system tied to a simple observe/attack/retreat program, I would imagine. The latter isn't very hard at all, speaking as someone who occasionally mods games to make the AI challenging, and the former can be bought off the shelf. Integrating it is probably the major issue, but I suspect someone who can throw together a manual drone could figure out a way beyond that.
Because of the ease of manufacture, drones would indeed be a perfect weapon for any group looking to reap the benefits of hi-tech, asymmetric warfare, including terrorist organizations. Creating a drone with a specified command to do a set amount of recon, then strike the largest cluster of people within a geographic zone would not be hard, and tracing the attack may well be impossible. Especially if the drone is the weapon. Of course, you could do that with a manual drone as well (Hezbollah used their own drones to carry out suicide attacks on the Israeli Navy), but an autonomous drone would allow for greater distance between the culprits and the targets, and would likely prevent all attempts at jamming signals that may allow a manual drone to be disabled.
Also, let's think a little more creatively here, like DARPA are doing. Those huge Predator drones are not the only type of drone the US military is looking to deploy. There has been considerable interest in insect size "assassinstion drones" for battlefield (and presumably off the battlefield counter-terrorism and insurgency) uses. And no doubt the land-based drones being worked on in various labs are being considered, for places where airborne drones would have difficulty in targeting a subject, such as built-up, urban areas.
Autonomous assassination drones, the size of an insect. Now there's a worrying thought.