vtek So can we say for 1000% that AI absolutely cannot act on its own/escape its boundaries without being given the latitude (syscalls for ex.) to do so?
Code, doesn't matter how complicated or complex it can be, still does only what has been coded to do.
If it helps you having a different point of view, you can see the word 'AI' as an umbrella term that describe a certain "pattern" in code, which has been covered in an very simplified way here:
raccoondad which is a very large function w/ in -> out and that's it
Speaking about Image recognition, if you dig back enough you will find that it started waaay before the advent of LLMs, which are just one kind of AI you can find.
These models are different just in how they determine what's in the picture, but as Foggy has covered already, with the image as the input, they will return a metrix as an output (with the example being cat=0.4% dog=22% sunset=9%).
This result is then given back to the caller (the Gallery app) which decides what to do next based on this output.
Unless in the Gallery app at a certain point you find something like if image.contains("cat"); then delete(image); fi (Code purposely not accurate), no data will be deleted, ever.
In other words, if the developer don't decide to delete files based on the ooutput of these models, no file will be deleted.
As said also by Tandara , it's what around the models that matters, because the models themselves are only giving back probabilities, basically.
The concept has been oversimplified of course, but I hope it makes a little bit clearer why "AIs" cannot go rogue.