[data-ottawa]

Does anyone know what's the deal with these scrapers, or why they're attributed to AI?

I would assume any halfway competent LLM driven scraper would see a mass of 404s and stop. If they're just collecting data to train LLMs, these seem like exceptionally poorly written and abusive scrapers written the normal way, but by more bad actors.

Are we seeing these scrapers using LLMs to bypass auth or run more sophisticated flows? I have not worked on bot detection the last few years, but it was very common for residential proxy based scrapers to hammer sites for years, so I'm wondering what's different.

[embedding-shape]

I just threw up a public Forjego instance for some lightweight collaboration. About 2 minutes after the certificate was created, I'm guessing they picked up the instance from the transparency logs for certificates, and started going through every commit and so on from the two repositories I had added.

Watched it for a while, thinking eventually it'd end. It didn't, seemed like Claudebot and GPTBot (which was the only two I saw, but could have been forged) went over the same URLs over and over again. They tried a bunch of search queries too at the same time.

The day after I got tired of seeing it so added a robot.txt forbidding any indexing. Waited a few hours, saw that they were still doing the same thing, so threw up basic authentication with `wiki:wiki` as the username:password basically, wrote the credentials on the page where I linked it and as expected they stopped trying after that.

They don't seem to try to bypass anything, whatever you put in front will basically defeat them except blocking them by user-agent, then they just switch to a browser-like user-agent instead, which is why I went the "trivial basic authentication" path instead.

Wasn't really an issue, just annoying when they try to masquerade as normal users. Had the same issue with a wiki instance, added rate limits and eventually they seemingly backed off more than my limits were set too, so I guess they eventually got it. Just checked the logs and seems they've stopped trying completely.

Seemingly it seems like people who are paying for their hosting by usage (which never made sense to me) is the ones hard hit by this. I'm hosting my stuff on a VPS, and don't understand what the big issue is, worst case scenario I'd add more aggressive caching and it wouldn't be an issue anymore.

[rozab]

I had the same issue when I first put up my gitea instance. The bots found the domain through cert registration in minutes, before there were any backlinks. GPTbot, ClaudeBot, PerplexityBot, and others.

I added a robots.txt with explicit UAs for known scrapers (they seem to ignore wildcards), and after a few days the traffic died down completely and I've had no problem since.

Git frontends are basically a tarpit so are uniquely vulnerable to this, but I wonder if these folks actually tried a good robots.txt? I know it's wrong that they ignore wildcards, but it does seem to solve the issue

[Imustaskforhelp]

Huh, I had a gitea instance in the public web on one of my netcup vps's. I didn't set any logs and was using cloudflare tunnels (with a custom bash script which makes cf tunnels expose PORT SUBDOMAIN).

Maybe its time for me to go ahead and start it again with logs to see if there are any logs.

I will maybe test it in all three 1) With CF tunnels + AI Block, 2) Only CF tunnels, 3) On a static IP directly. Maybe you can try the experiment too and we can compare our findings (also saying because I am lazy and I had misconfigured that cf tunnel so when it quit, I was too lazy to restart the vps given I just use it as a playground and just wanted to play around self hosting but maybe I will do it again now)

[Fabricio20]

Since you had the logs for this, can you confirm the IP ranges they were operating from? You mention "Claudebot and GPTBot" but I'm guessing this is based off of the user-agent presented by the scrapers and could easily be faked to shift blame. I genuinely doubt Anthropic and such would be running scrapers that are this badly written/implemented, it doesnt make economic sense. I'd love to see some of the web logs from this if you'd be willing to share! I feel like this is just some of the old scraper bots now advertising themselves as AI bots to shift blame into the AI companies.

[trumpdong]

One idea is to not block them, but return a plausible looking page that's not the normal page, then it won't be detected as a block.

[simonw]

I would love to understand this.

Just a few years ago badly behaved scrapers were rare enough not to be worth worrying about. Today they are such a menace that hooking any dynamic site up to a pay-to-scale hosting platform like Vercel or Cloud Run can trigger terrifying bills on very short notice.

"It's for AI" feels like lazy reasoning for me... but what IS it for?

One guess: maybe there's enough of a market now for buying freshly updated scrapes of the web that it's worth a bunch of chancers running a scrape. But who are the customers?

[SCHiM]

The bar to ingest unstructured data into something usable was lowered, causing more people to start doing it.

Used to be you needed to implement some papers to do sentiment analysis. Reasonably high bar to entry. Now anyone can do it, the result: more people doing scraping (in less competent scrapers too).

[devsda]

For whatever reason, legislation is lax right now if you claim the purpose of scraping is for AI training even for copyrighted material.

May be everyone is trying to take advantage of the situation before law eventually catches up.

[ben_w]

I'm currently on a free trial of ChatGPT, and one new thing I didn't regularly see before* is that longer tasks perform a lot of web searches when generating non-trivial results.

I wonder if this is part of it? It's not (just) DDOS by crawlers, it's DDOS by the users themselves triggering (albeit indirectly) far more requests than a human normally would? I've seen that happen in a different context, over a decade ago now.

* old models would do this sometimes when you ask for whatever the "deep research" mode was called, but this now seems to happen a lot more and involve a lot more fetches

[zerocrates]

I would say there's a couple aspects.

The crawlers for the big famous names in AI are all less well behaved and more voracious than say, Googlebot. Though this is all somewhat muddied by companies that ran the former "good" crawlers all also being in the AI business and sometimes trying to piggyback on people having allowed or whitelisted their search crawling User-Agent, mostly this has settled a little where they're separating Googlebot from GoogleOther, facebookexternalhit from meta-externalagent, etc. This was an earlier "wave" of increased crawling that was obviously attributable to AI development. In some cases it's still problematic but this is generally more manageable.

The other stuff, the ones that are using every User-Agent under the sun and a zillion datacenter IPs and residential IPs and rotate their requests constantly so all your naive and formerly-ok rate-based blocking is useless... that stuff is definitely being tagged as "for AI" on the basis of circumstantial evidence. But from the timing of when it seemed to start, the amount of traffic and addresses, I don't have any problem guessing with pretty high confidence that this is AI. To your question of "who are the customers"... who's got all the money in the world sloshing around at their fingertips and could use a whole bunch of scraped pages about ~everything? Call it lazy reasoning if you'd like.

How much this traces back ultimately to the big familiar brand names vs. would-be upstarts, I don't know. But a lot of sites are blocking their crawlers that admit who they are, so would I be surprised to see that they're also paying some shady subcontractors for scrapes and don't particularly care about the methods? Not really.

[Tharre]

> Does anyone know what's the deal with these scrapers, or why they're attributed to AI?

You don't really need to guess, it's obvious from the access logs. I realize not everyone runs their own server, so here are a couple excerpts from mine to illustrate:

- "meta-externalagent/1.1 +https://developers.facebook.com/docs/sharing/webmasters/craw...)"

- "Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; ClaudeBot/1.0; [e-mail])"

- "Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; Amazonbot/0.1; +https://developer.amazon.com/support/amazonbot) Chrome/119.0.6045.214 Safari/537.36"

- "Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; GPTBot/1.3; +https://openai.com/gptbot)"

- [...] (compatible; PetalBot;+https://webmaster.petalsearch.com/site/petalbot)"

And to give a sense of scale, my cgit instance recieved 37 212 377 requests over the last 60 days, >99% of which are bots. The access.log from nginx grew to 12 GiB in those 60 days. They scrape everything they can find, indiscriminately, including endpoints that have to do quite a bit of work, leading to a baseline 30-50% CPU utilization on that server right now.

Oh, and of course, almost nothing of what they are scraping actually changed in the last 60 days, it's literally just a pointless waste of compute and bandwidth. I'm actually surprised that the hosting companies haven't blocked all of them yet, this has to increase their energy bills substantially.

Some bots also seem better behaved then others, OpenAI alone accounts for 26 million of those 37 million requests.

[everybodyknows]

Following your link above, https://openai.com/gptbot

> ChatGPT-User is not used for crawling the web in an automatic fashion. Because these actions are initiated by a user, robots.txt rules may not apply.

So, not AI training in this case, nor any other large-batch scraping, but rather inference-time Retrieval Augmented Generation, with the "retrieval" happening over the web?

[themafia]

There's value to be had in ripping the copyright off your stuff so someone else can pass it off as their stuff. LLMs have no technical improvements so all they can do is throw more and more stolen data into it and hope it, somehow, crosses a nebulous "threshold" where it suddenly becomes actually profitable to use and sell.

It's a race to the bottom. What's different is we're much closer to the bottom now.

[jillesvangurp]

Using an LLM to ponder responses for requests is way too costly and slow. Much easier to just use the shotgun approach and fire off a lot of requests and deal with whatever bothers to respond.

This btw is nothing new. Way back when I still used wordpress, it was quite common to see your server logs filling up with bots trying to access endpoints for commonly compromised php thingies. Probably still a thing but I don't spend a lot of time looking at logs. If you run a public server, dealing with maliciously intended but relatively harmless requests like that is just what you have to do. Stuff like that is as old as running stuff on public ports is.

And the offending parties writing sloppy code that barely works is also nothing new.

AI opportunism certainly has added a bit of opportunistic bot and scraper traffic but it doesn't actually change the basic threat model in any fundamental way. Previously version control servers were relatively low value things to scrape. But code just became interesting for LLMs to train on.

Anyway, having any kind of thing responding on any port just invites opportunistic attempts to poke around. Anything that can be abused for DOS purposes might get abused for exactly that. If you don't like that, don't run stuff on public servers or protect them properly. Yes this is annoying and not necessarily easy. Cloud based services exist that take some of that pain away.

Logs filling up with 404, 401, or 400 responses should not kill your server. You might want to implement some logic that tells repeat offenders 429 (too many requests). A bit heavy handed but why not. But if you are going to run something that could be used to DOS your server, don't be surprised if somebody does that.

[arnarbi]

> why they're attributed to AI?

I don’t think they mean scrapers necessarily driven by LLMs, but scrapers collecting data to train LLMs.

[danaris]

> If they're just collecting data to train LLMs, these seem like exceptionally poorly written and abusive scrapers written the normal way, but by more bad actors.

Right, this is exactly what they are.

They're written by people who a) think they have a right to every piece of data out there, b) don't have time (or shouldn't have to bother spending time) to learn any kind of specifics of any given site and c) don't care what damage they do to anyone else as they get the data they crave.

(a) means that if you have a robots.txt, they will deliberately ignore it, even if it's structured to allow their bots to scrape all the data more efficiently. Even if you have an API, following it would require them to pay attention to your site specifically, so by (b), they will ignore that too—but they also ignore it because they are essentially treating the entire process as an adversarial one, where the people who hold the data are actively trying to hide it from them.

Now, of course, this is all purely based on my observations of their behavior. It is possible that they are, in fact, just dumb as a box of rocks...and also don't care what damage they do. (c) is clearly true regardless of other specific motives.

[hsuduebc2]

I’m guessing, but I think a big portion of AI requests now come from agents pulling data specifically to answer a user’s question. I don’t think that data is collected mainly for training now but are mostly retrieved and fed into LLMs so they can generate the response. Thus so many repeated requests.

[everforward]

I think it’s a) volume of scrapers, and b) desire for _all_ content instead of particular content, and c) the scrapers are new and don’t have the decades of patches Googlebot et al do.

5 years ago there were few people with an active interest in scraping ForgeJo instances and personal blogs. Now there are a bajillion companies and individuals getting data to train a model or throw in RAG or whatever.

Having a better scraper means more data, which means a better model (handwavily) so it’s a competitive advantage. And writing a good, well-behaved distributed scraper is non-trivial.

[dirkc]

I'm hazarding a guess that there are many AI startups that focus on building datasets with the aim to sell those datasets. Still doesn't make total sense, since doing it badly would only hurt them, but maybe they don't really care about the product / outcome, they're just capturing their bit of the AI goldrush?

[sh34r]

I guess the AI companies finally figured out they’re supposed to buy their stolen datasets from a shell company spun up by the most unsavory character within two degrees of the CEO. Every CEO has a drug dealer, and every CEO drug dealer knows the greasy grey hat dude running a data laundry “startup.” The VCs usually know some private equity dons who run the same racket to do bust out fraud, too.

It’s truly unbelievable that OpenAI and Anthropic were so sloppy. Pirating all that copyrighted media and not even bothering to hide behind one layer of indirection. Amateurs.

So yeah… it’s what, five years’ worth of pent up demand for organized crime, hitting the market everywhere all at once? I’m surprised the request volume isn’t higher!

[golem14]

You can always prune it down later (so the thinking goes, no doubt)

[M95D]

I stopped trying to understand. Encountering a 404 on my site leads directly to a 1 year ban.

[embedding-shape]

Damn, as someone who sometimes navigate by guessing URLs and rewriting them manually in the address bar, I hope more don't start doing this, I probably see at least one self-inflicted 404 per day at least.

[octoberfranklin]

They're rotating through huge pools of residential IP addresses.

[tasuki]

Sounds like you're keeping all your URLs alive forever? Commendable!

[octoberfranklin]

I don't think it has anything to do with LLMs.

I think the big cloud companies (AWS) figured out that they could scrape compute-intensive pages in order to drive up their customers' spend. Getting hammered? Upgrade to more-expensive instances. Not using cloud yet? We'll force you to.

The other possibility is cloudflare punishing anybody who isn't using it.

Probably a combination of these two things. Whoever's behind this has ungodly supplies of cheap bandwidth -- more than any AI company does. It's a cloud company.

[yellowapple]

> Whoever's behind this has ungodly supplies of cheap bandwidth -- more than any AI company does. It's a cloud company.

Most of the major cloud companies are themselves also AI companies, so I don't think the “cloud companies are artificially driving up compute spend” hypothesis is mutually exclusive with the “AI companies are doing a very bad job at scraping” hypothesis.