Until drones deploy counter-blinkenlights. As someone who has built a realtime people tracker art installation in a disco: Stroboscopes are highly effective at confusing these models.
A military system is being launched at an area to destroy a target that is presumed to be trying to hide (hence AI).
Absolutely nothing stops a secondary, much simpler identifier system of "home on bright flickering lights".
Any system where you try to confuse an image recognition model by loudly doing something unexpected suffers from the fact that regular tracking systems have been good at finding those for a long time.
No AI model confusion tech is worthwhile if it otherwise makes something much more visible by other means - it's a scifi trope that people want something "obvious" to be invisible to the machine. It's not how reality works though.
There's a limit to the number of targeting systems you can afford to deploy on a weapon, and there's the problem of reconciling their signals if their views of where the target is start to diverge.
"Counter-blinkenlights" are well-known as thermal flares, and more similar but as crude, [dazzlers] that blind the incoming rockets' sensors.
AFAIU, a dazzler is intended to blind the whole sensor, so home on dazzler would need specialized hardware, not only a software adjustment. (Replying to other messages in this thread)
But it's not a separate system, it's just a separate algorithm - it's software.
"No target" by AI but "target" by the "engage light sources" recognizer is still successful targeting.
"No target" versus "aim at anything not green or blue" is as well.
It's also a question of the speeds involved: a dazzler on high speed jet versus a high speed missile might cause a near miss which means hundreds of meters of actual separation. Diving drone vs. truck? Not so much.
Ya, things always evolve, and can't always be perfect, especially against adapting enemies. Strobes; ya - would be interesting to see what these do vs us, not tried.