LLM training bots are a plague
LLM training bots are a plague
Excerpt from a message I just posted in a #diaspora team internal forum category. The context here is that I recently get pinged by slowness/load spikes on the diaspora* project web infrastructure (Discourse, Wiki, the project website, ...), and looking at the traffic logs makes me impressively angr...
Evidence for the DDoS attack that bigtech LLM scrapers actually are.
You're viewing a single thread.
How feasible is it to configure my server to essentially perform a reverse-slow-lorris attack on these LLM bots?
If they won't play nice, then we need to reflect their behavior back onto themselves.
Or perhaps serve a 404, 304 or some other legitimate-looking static response that minimizes load on my server whilst giving then least amount of data to train on.
The only simple possibles ways are:
- robot.txt
- rate limiting by ip
- blocking by user agent
From the article, they try to bypass all of them:
They also don't give a single flying fuck about robots.txt ...
If you try to rate-limit them, they'll just switch to other IPs all the time. If you try to block them by User Agent string, they'll just switch to a non-bot UA string (no, really). This is literally a DDoS on the entire internet.
It then become a game of whac a mole with big tech 😓
The more infuriating for me is that it's done by the big names, and not some random startup.Edit: Now that I think about it, this doesn't prove it is done by Google or Amazon: it can be someone using random popular user agentsI do believe there's blocklists for their IPs out there, that should mitigate things a little
A possibility to game this kind of bots is to add a hidden link to a randomly generated page, which contain itself a link to another random page, and so one.: The bots will still consume resources but will be stuck parsing random garbage indefinitely.
I know there is a website that is doing that, but I forget his name.Edit: This is not the one I had in mind, but I find https://www.fleiner.com/bots/ describes be a good honeypot.
maybe you mean this incident https://news.ycombinator.com/item?id=40001971
This is it, thanks: https://www.web.sp.am/
Should serve an alternate site that is not quite right to those user agents