3 ms·
Slightly unrelated… This reads like yet another major attack vector through LLMs. A threat actor only needs to provide a few mentions of the package on reputab
by skilled 2y ago
Slightly unrelated…
This reads like yet another major attack vector through LLMs. A threat actor only needs to provide a few mentions of the package on reputable sites and it is almost guaranteed that your backdoored package will be mentioned by the LLM once it is retrained with new data, or even in cases like RAG.
I am guessing I am not the first to think of this though. Wouldn’t be surprised if this kind of attack vector is already being set in motion for all kinds of other purposes; product reviews, etc.
- vertis 2y agoIt is an important consideration for live language models. The example recently of a bad Reddit answer popping up in Gemini -- putting glue on the pizza demonstrates how easy it is. While there are definitely ethical issues with the data that they are using. Trainers need to get a handle on this kind of thing because for LLMs to be truly useful they have to consume large parts of the web.
- nottorp 2y agoPerhaps a ... human curated index ... would be more useful then? :) I mean, if you're manually curating what you feed to the LLM you end up doing one of the web directories of old and might as well skip the LLM part... or use it just to gain funding...