Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

It's hard to gauge the effectiveness of poisoning huge training sets since anything you do is a figurative drop in the ocean, but if you can poison the small amount of data that an AI agent requests on-the-fly to use with RAG then I would guess it's much easier to derail it.


This study shows that controlling 0.1% may be enough.

https://arxiv.org/abs/2410.13722v1

I have noticed some popular copied but incorrect leetcode examples leaking into the dataset.

I suspect it depends on domain specificity, but that seems within the ability of an SEO spammer or decentralized group of individuals.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: