What I meant was that a specialized tool could be built with an LLM backend that analyzed the code for what kind of output, if any, it created. We know already that it can do that because you've written about it and so have I. Surely it could do this work faster than people and find many of those spam/garbage repo cases.
That’s like arguing against using locks on doors because they’re pickable.
You’re right: they can be defeated.
But they might cut it by 80-90%, and be complemented with other tools to reduce the flood to a trickle.
The problem with those real-world analogies is that those things don't scale in the real world. Even if you're a 10x lockpicker compared to an average burglar, you still have to actually go to the place you want to steal from, actually carry out the loot, expose yourself to being witnessed, and all that stuff.
Whereas with computers, if you have, say, a zero-day exploit for nginx, it's feasible for a small band of black hats to infect hundreds of thousands of servers. And if a single person has the equivalent of a zero-day exploit for NPM's hypothetical review AI, they can just spam tens of thousands of modules and if only 0.1% manage to slip through the cracks, you're golden.
Fresh story this evening says it's already being used for that.
https://www.theregister.com/2023/03/30/socket_chatgpt_malwar...
The AI that uses LLM as a component, rather than consisting of an LLM, would be harder to fool, but we don't have that yet, despite the way we keep pretending that LLMs are already that.