Getting a dejavu of a couple SF books where a party needed to obtain totally-not-weapons, sometimes even including talking an AI into building them (with sometimes the AI even being on it basically, just not being able to acknowledge it due to outside setup filters even!).
Laser flashlights and rapid rescue shuttles come to mind. ;-)
Makes me think of the sentient hell-class weapons in the Revelation Space series that (if I recall correctly) had to be goaded and convinced into firing
I think you are pretty close. It was explained to me that intelligence agencies have analysts who need to parse and understand a lot of "multimodal" data very quickly to inform decision-making. Maybe the thinking is that LLMs could help with it? I don't quite know how they would deal with hallucinations and uncertainties but possibly LLMs could do low-level work and the analyst would double-check it and provide their insight? Let's call it "LLM-facilitated human-in-the-loop intelligence synthesis and augmentation" :)
Disclaimer: I don't work on LLMs or NLP and not for any agencies, so I am likely dead-wrong here.
Take a look at what you can find about Google’s Project Nimbus. It’s my understanding from public info that this system is active in Israel today, and from day one of the war was doing multimodal analysis of the content posted online by Hamas to locate the hostages in some automated fashion.
"To improve our response to emerging threats, our SmartMissiles™ now use OpenAI for fire/don't-fire decisions instead of a human operator."