Australia already has area-denial sentries that autonomously shoot at any motion (with some parameters regarding size and speed).

These, or a similar techology was used for a while along the Korean DMZ until we started talking about building autonomous drones.

One of the shot-down airliner incidents (Flight 007, maybe?) involved a misdesignation of a sensor contact by a US Aegis missileboat system. The plane was pinging with F4 Phantom Radar (which ruled out an ordinary airliner). The Aegis required a human to authorize an attack, but it reported the contact as a bogey (unknown, peresumed to be hostile)...

— Apparently, I posted this without finishing it. —

So that instance might be considered the first historical case of an autonomous weapon system accidentally killing a civilian (at least partially civilan) target, given the human doing the authorizing had inadequate data to make an informed decision.

(A lot of cruelty of our systems comes from authorizations based on partial data. Law enforcement in the US is renowned for massaging their warrants to make them easy on the signing magistrate, resulting in kids and dogs slain during SWAT raids in poor neighborhoods. I'm ranting.)