Codeberg: army of AI crawlers are extremely slowing us; AI crawlers learned how to solve the Anubis challenges.
cross-posted from: https://programming.dev/post/35852706
There once was a dream of the semantic web, also known as web2. The semantic web could have enabled easy to ingest information of webpages, removing soo much of the computation required to get the information. Thus preventing much of the air crawling cpu overhead.
What we got as web2 instead was social media. Destroying facts and making people depressed at a newer before seen rate.
Web3 was about enabling us to securely transfer value between people digitally and without middlemen.
What crypto gave us was fraud, expensive jpgs and scams. The term web is now even so eroded that it has lost much of its meaning. The information age gave way for the misinformation age, where everything is fake.
Capitalism is grand, innit. Wait, not grand, I meant to say cancer
I feel like half of the blame capitalism gets is valid, but the other half is just society. I don't care what kind of system you're under, you're going to have to deal with other people.
Oh, and if you try the system where you don't have to deal with people, that just means other people end up handling you.
Web3 was about enabling us to securely transfer value between people digitally and without middlemen.
It's ironic that the middlemen showed up anyway and busted all the security of those transfers
You want some bipcoin to buy weed drugs on the slip road? Don't bother figuring out how to set up that wallet shit, come to our nifty token exchange where you can buy and sell all kinds of bipcoins
oh btw every government on the planet showed up and dug through our insecure records. hope you weren't actually buying shroom drugs on the slip rod
also we got hacked, you lost all your bipcoins sorry
At least, that's my recollection of events. I was getting my illegal narcotics the old fashioned way.
also we got hacked, you lost all your bipcoins sorry
aaaaaaaaand - it's gone!
You want some bipcoin to buy weed drugs on the slip road? Don’t bother figuring out how to set up that wallet shit, come to our nifty token exchange where you can buy and sell all kinds of bipcoins
Maybe I'm slow today, but what is this referencing? Most dark web sites use Monero. Is there some centralized token that people used instead?
Edit: Oh, I guess you're referring to Mt.Gox? I mean yeah, people were pretty stupid for keeping their bitcoin in exchange wallets (and sending it right to the drug dealers directly from there? Real dumb). That's always a bad idea. I don't think they transferred it there instead of something else, they just never took custody of the coins after buying them on the exchange.
Sound like it went the same way everything else went. The less money is involved the more trustworthy it is.
Mr. Internet, tear down these walls! (for all these walled gardens)
Return the internet to the wild. Let it run feral like dinosaurs on an island.
Let the grannies and idiots stick themselves in the reservations and asylums run by billionaires.
Let's all make Neocities pages about our hobbies and dirtiest, innermost thoughts. With gifs all over.
Web3 was about enabling us to securely transfer value between people digitally and without middlemen
I don't think it ever was that, I think folding ideas has the best explanation of what it was meant to be, it was meant to be a way to grab power, away from those who already have it
Much drama.
I agree about semantic web, but the issue is with all of the Internet. Both its monopoly as the medium of communication, and its architecture.
And if we go semantic for webpages, allowing the clients to construct representation, then we can go further, to separate data from medium, making messages and identities exist in a global space, as they (sort of, need a better solution) do in Usenet.
About the Internet itself being the problem - that's because it's hierarchical, despite appearances, and nobody understands it well. Especially since new systems of this kind are not being built often, to say the least, so the majority of people using the Internet doesn't even think about it as a system. It takes it for given that this is the only paradigm for the global network. And that it's application-neutral, which may not be true.
20 years ago, when I was a kid, people would think and imagine all kinds of things about the Internet and about the future and about ways all this can break, and these were normal people, not tech types, and one would think with time we wouldn't become more certain, as it becomes bigger and bigger.
OK, I'm just having an overvalued idea that the Internet is poisoned. Bad sleep, nasty weather, too much sweets eaten. Maybe that movement of packets on the IP protocol can somehow give someone free computation, with enough machines under their control, by using counters in the network stack as registers, or maybe something else.
Preach!
Anubis isn't supposed to be hard to avoid, but expensive to avoid. Not really surprised that a big company might be willing to throw a bunch of cash at it.
This is what I've kept saying about POW being a shit bot management tactic. Its a flat tax across all users, real or fake. The fake users are making money to access your site and will just eat the added expense. You can raise the tax to cost more than what your data is worth to them, but that also affects your real users. Nothing about Anubis even attempts to differentiate between bots and real users.
If the bots take the time, they can set up a pipeline to solve Anubis tokens outside of the browser more efficiently than real users.
No, it's expensive to comply (at a massive scale), but easy to avoid. Just change the user agent. There's even a dedicated extension for bypassing Anubis.
Even then AI servers have plenty of compute, it realistically doesn’t cost much. Maybe like a thousandth of a cent per solve? They're spending billions on GPU power, they don't care.
I've been saying this since day 1 of Anubis but nobody wants to hear it.
I feel like at some point it needs to be active response. Phase 1 is a teergrube type of slowness to muck up the crawlers, with warnings in the headers and response body, and then phase 2 is a DDOS in response or maybe just a drone strike and cut out the middleman. Once you've actively evading Anubis, fuckin' game on.
I think the best thing to do is to not block them when they're detected but poison them instead. Feed them tons of text generated by tiny old language models, it's harder to detect and also messes up their training and makes the models less reliable. Of course you would want to do that on a separate server so it doesn't slow down real users, but you probably don't need much power since the scrapers probably don't really care about the speed
I love catching bots in tarpits, it's actually quite fun
Some guy also used zip bombs against AI crawlers, don't know if it still works. Link to the lemmy post
Yeah that was my thought. Don't reject them, that's obvious and they'll work around it. Feed them shit data - but not too obviously shit - and they'll not only swallow it but eventually build up to levels where it compromises them.
I've suggested the same for plain old non-AI data stealing. Make the data useless to them and cost more work to separate good from bad, and they'll eventually either sod off or die.
A low power AI actually seems like a good way to generate a ton of believable - but bad - data that can be used to fight the bad AI's. It doesn't need to be done real-time either as datasets can be generated in advance
Wasn't this called black ice in Neuromancer? Security systems that actively tried to harm the hacker?
These crawlers come from random people’s devices via shady apps. Each request comes from a different IP
Most of these AI crawlers are from major corporations operating out of datacenters with known IP ranges, which is why they do IP range blocks. That's why in Codeberg's response, they mention that after they fixed the configuration issue that only blocked those IP ranges on non-Anubis routes, the crawling stopped.
For example, OpenAI publishes a list of IP ranges that their crawlers can come from, and also displays user agents for each bot.
Perplexity also publishes IP ranges, but Cloudflare later found them bypassing no-crawl directives with undeclared crawlers. They did use different IPs, but not from "shady apps." Instead, they would simply rotate ASNs, and request a new IP.
The reason they do this is because it is still legal for them to do so. Rotating ASNs and IPs within that ASN is not a crime. However, maliciously utilizing apps installed on people's devices to route network traffic they're unaware of is. It also carries much higher latency, and could even allow for man-in-the-middle attacks, which they clearly don't want.
Or your TV or IOT devices. Residential proxies are extremely shady businesses.
Yep
Yes. A nonprofit organization in Germany is going to be launching drone strikes globally. That is totally a better world.
Its also important to understand that a significant chunk of these botnets are just normal people with viruses/compromised machines. And the fastest way to launch a DDOS attack is to... rent the same botnet from the same blackhat org to attack itself. And while that would be funny, I would also rather orgs I donate to not giving that money to blackhat orgs. But that is just me.
Right
Tech bros just actively making the internet worse for everyone.
Tech bros just actively making
the internetsociety worse for everyone.
FTFY.
I mean, tech bros of the past invented the internet
Those are not the tech bros. The tech bros are the ones who move fast and break things. The internet was built by engineers and developers
Nah, that was DARPA
Those were tech nerds. "Tech bros" are jabronis who see the tech sector as a way to increase the value of the money their daddies gave them.
I mean, we really have to ask ourselves - as a civilization - whether human collaboration is more important than AI data harvesting.
I think every company in the world is telling everyone for a few months now that what matter is AI data harvesting. There's not even a hint of it being a question. You either accept the AI overlords or get out of the internet. Our ONLY purpose it to feed the machine, anything else is irrelevant. Play along or you shall be removed.
We need to poison better.
get out of the internet.
At some point, this would be the best option, sadly
I was fine before the AI.
The biggest customer of AI are the billionaires who can't hire enough people for their technofeudalist/surveillance capitalism agenda. The billionaires (wannabe aristocrats) know that machines have no morals, no bottom lines, no scruples, don't leak info to the press, don't complain, don't demand to take time off or to work from home, etc.
AI makes the perfect fascist.
They sell AI like it's a benefit to us all, but it ain't that. It's a benefit to the billionaires who think they own our world.
AI is used for censorship, surveillance pricing, activism/protest analysis, making firing decisions, making kill decisions in battle, etc. It's a nightmare fuel under our system of absurd wealth concentration.
Fuck AI.
Can there be a challenge that actually does some maliciously useful compute? Like make their crawlers mine bitcoin or something.
Did you just say use the words "useful" and "bitcoin" in the same sentence? o_O
Not without making real users also mine bitcoin/avoiding the site because their performance tanked.
The Monero community spent a long time trying to find a "useful PoW" function. The problem is that most computations that are useful are not also easy to verify as correct. javascript optimization was one direction that got pursued pretty far.
But at the end of the day, a crypto that actually intends to withstand attacks from major governments requires a system that is decentralized, trustless, and verifiable, and the only solutions that have been found to date involve algorithms for which a GPU or even custom ASIC confers no significant advantage over a consumer-grade CPU.
Anubis does that (the computation part). You may've seen it already.
Crazy. DDoS attacks are illegal here in the UK.
So, sue the attackers?
The problem is that hundreds of bad actors doing the same thing independently of one another means it does not qualify as a DDoS attack. Maybe it’s time we start legally restricting bots and crawlers, though.
I know this is the most ridiculous idea, but we need to pack our bags and make a new internet protocol, to separate us from the rest, at least for a while. Either way, most “modern” internet things (looking at you, JavaScript) are not modern at all, and starting over might help more than any of us could imagine.
Like Gemini?
From official Website:
Gemini is a new internet technology supporting an electronic library of interconnected text documents. That's not a new idea, but it's not old fashioned either. It's timeless, and deserves tools which treat it as a first class concept, not a vestigial corner case. Gemini isn't about innovation or disruption, it's about providing some respite for those who feel the internet has been disrupted enough already. We're not out to change the world or destroy other technologies. We are out to build a lightweight online space where documents are just documents, in the interests of every reader's privacy, attention and bandwidth.
Won't the bots just adapt and move there too?
It's not the most well thought-out, from a technical perspective, but it's pretty damn cool. Gemini pods are a freakin' rabbi hole.
I've personally played with Gemini a few months ago, and now want a new Internet as opposed to a new Web.
Replace IP protocols with something better. With some kind of relative addressing, and delay-tolerant synchronization being preferred to real-time connections between two computers. So that there were no permanent global addresses at all, and no centralized DNS.
With the main "Web" over that being just replicated posts with tags hyperlinked by IDs, with IDs determined by content. Structured, like semantic web, so that a program could easily use such a post as directory of other posts or a source of text or retrieve binary content.
With user identities being a kind of post content, and post authorship being too a kind of post content or maybe tag content, cryptographically signed.
Except that would require to resolve post dependencies and retrieve them too with some depth limit, not just the post one currently opens, because, if it'd be like with bittorrent, half the hyperlinks in found posts would soon become dead, and also user identities would possibly soon become dead, making authorship check impossible.
And posts (suppose even sites of that flatweb) being found by tags, maybe by author tag, maybe by some "channel" tag, maybe by "name" tag, one can imagine plenty of things.
The main thing is to replace "clients connecting to a service" with "persons operating on messages replicated on the network", with networked computers sharing data like echo or ripples on the water. In what would be the general application layer for such a system.
OK, this is very complex to do and probably stupid.
It's also not exactly the same level as IP protocols, so this can work over the Internet, just like the Internet worked just fine, for some people, over packet radio and UUCP or FTN email gates and copper landlines. Just for the Internet to be the main layer in terms of which we find services, on the IP protocols, TCP, UDP, ICMP, all that, and various ones and DNS on application layer, - that I consider wrong, it's too hierarchical. So it's not a "replacement".
Gosh. Corporations are rampantly attempting to access resources so they can perform copyright infringement en-masse. I wonder if there is a legal mechanism to stop them? Oh, no there isn't because our government is fully corrupted.
I think, in this particular case, it's aggressive apathy/incompetence and not malice. Remember, Trump didn't even know what Nvidia was.
AI's don't have a skin color or use the bathroom so you can't whip your cult into a frenzy by Othering it. You can't solidify your fascism by getting bogged down in the details of IP law.
We had a trust based system for so long. No one is forced to honor robots.txt, but most big players did. Almost restores my faith in humanity a little bit. And then AI companies came and destroyed everything. This is why we can't have nice things.
Big players are the ones behind most AIs though.
Is there nightshade but for text and code? Maybe my source headers should include a bunch of special characters that then give a prompt injection. And sprinkle some nonsensical code comments before the real code comment.
I think the issue is that text uses comparatively very little information, so you can't just inject invisible changes by changing the least insignificant bits - you'd need to change the actual phrasing/spelling of your text/code, and that'd be noticable.
There are glitch tokens but I think those only effect it when using it.
It's always a cat-n-mouse game.
Except previously bombarding another person's server for personal gain was illegal.
I don't know if this is news to you, but most of the internet never cared about what's legal or not.
Not if it's AI.
/s aside, maybe you could call'em out on involuntary DoSing, but then slashdot and similar sites would get into trouble.
reminder to donate to codeberg and forgejo :)
Eventually we'll have "defensive" and "offensive" llm's managing all kinds of electronic warfare automatically, effectively nullifying each other.
That's actually a major plot point in Cyberpunk 2077. There's thousands of rogue AI's on the net that are constantly bombarding a giant firewall protecting the main net and everything connected to it from being taken over by the AI.
Places like cloudflare and akamai are already using machine learning algorithms to detect bot traffic at a network level. You need to use similar machine learning to evade them. And since most of these scrapers are for AI companies I'd expect a lot of the scrapers to be LLM generated.
Obligatory AI ≠ LLM. How would scrapers benefit from the LLMs they help train? The defense is obvious, LLM-generated slop traps against scrapers already exist.
Is there a migration tool? If not would be awesome to migrate everything including issues and stuff. Bet even more people would move.
Codeberg has very good migration tools built in. You need to do one repo at a time, but it can move issues, releases, and everything.
There are migration tools, but not a good bulk one that I could find. It worked for my repos except for my unreal engine fork.
Okay what about...what about uhhh... Static site builders that render the whole page out as an image map, making it visible for humans but useless for crawlers 🤔🤔🤔
Accessibility gets throw out the window?
I wasn't being totally serious, but also, I do think that while accessibility concerns come from a good place, there is some practical limitation that must be accepted when building fringe and counter-cultural things. Like, my hidden rebel base can't have a wheelchair accessible ramp at the entrance, because then my base isn't hidden anymore. It sucks that some solutions can't work for everyone, but if we just throw them out because it won't work for 5% of people, we end up with nothing. I'd rather have a solution that works for 95% of people than no solution at all. I'm not saying that people who use screen readers are second-class citizens. If crawlers were vision-based then I might suggest matching text to background colors so that only screen readers work to understand the site. Because something that works for 5% of people is also better than no solution at all. We need to tolerate having imperfect first attempts and understand that more sophisticated infrastructure comes later.
But yes my image map idea is pretty much a joke nonetheless
AI is pretty good at OCR now. I think that would just make it worse for humans while making very little difference to the AI.
Humans that don't see:
AI these days reads text from images better than humans can
Computer vision models can read/parse pixel geometry.
I'm ashamed to say that I switched my DNS nameservers to CF just for their anti crawler service.
Knowing Cloudflare, god know how much longer it'll be free for.
Did you enable the AI black hole/tarpit? It's the main reason I've used their stuff.
TIL! Just enabled it, thanks
I use Anubis on my personal website, not because I think anything I’ve written is important enough that companies would want to scrape it, but as a “fuck you” to those companies regardless
That the bots are learning to get around it is disheartening, Anubis was a pain to setup and get running
Question: those artificial stupidity bots want to steal the issues or want to steal the code? Because why they're wasting a lot of resources scraping millions of pages when they can steal everything via SSH (once a month, not 120 times a second)
That would require having someone with real intelligence running the scraper.
they just want all text
For mbin I managed to kill the attack of the scrapers only using cloudflare managed challenge for all except to fediverse post endpoints, from fediverse ua agents on certain get endpoints. Managed challenge on everything else.
So far, they've not gotten past it. But, a matter of time.
man, you'd think they'd just use the actual activitypub protocol to inhale all that data at once and not bother with costly scraping.
This A aint very I
AI was never intelligent. It’s a marketing term, that’s all. It has absolutely no meaning.
Same for all the WordPress blogs, by default in all of them there's an API without authentication that lets you download ALL the posts in an easy JSON.
Dear artificial stupidity bot... WHY THE FUCK ARE YOU FUCKING SCRAPING THE WHOLE PAGE 50 TIMES A SECOND????
Well the posts to inbox are generally for incoming info. Yes, there's endpoints for fetching objects. But, they don't work for indexing, at least not on mbin/kbin. If you have a link, you can use activitypub to traverse upwards from that object to the root post. But you cannot iterate down to child comments from any point.
The purpose is that say I receive an "event" from your instance. You click like on a post I don't have on my instance. Then the like event has a link to the object for that on activitypub. If I fetch that object it will have a link to the comment, if I fetch the comment it will have the comment it was in reply to, or the post. It's not intended to be used to backfill.
So they do it the old fashioned way, traversing the human side links. Which is essentially what I lock down with the managed challenge. And this is all on the free tier too.
Increasingly, I’m reminded of this: Paul Bunyan vs. the spam bot (or how Paul Bunyan triggered the singularity to win a bet). It’s a medium-length read from the old internet, but fun.
I knew that was the worse option. Use the one that traps them in an infinite maze.
You need to properly detect that they're bots first and then they'll just figure out how to spoof that. Then you're back to square one.
Abstractly, POW doesn't need to determine if you're a bot or not. To make a request, as a human or bot, you need to pay in cpu-time. The hope is that the cost is not so high that a human notices very much but for a bot trying to hoover up data as fast as possible, the aggregate cost is high.
I think the more horrifying aspect is that they'll just build ever bigger datacenters to crunch POW tests faster and the carbon cost will skyrocket even more.
Exactly. Imagine needing to pay a penny for every request. Not a huge deal for someone who only makes one or two requests per year. But if you’re running a bot farm and making tens of millions of requests per day, you’ll quickly find that your operating costs have skyrocketed. That’s basically the idea behind Anubis; Make someone pay in CPU time, so the legit users don’t really notice but bots quickly eat up all of their servers’ CPU.
Trap users in the maze as well :)
Oh I haven't even considered the carbon aspect. Anubis is an even worse idea than I previously thought...
Just provide a full dump.zip plus incremental daily dumps and they won't have to scrape ?
Isn't that an obvious solution ? I mean, it's public data, it's out there, do you want it public or not ?
Do you want it only on openai and google but nowhere else ? If so then good luck with the piranhas
The Wikimedia Foundation does just that, and still, their infrastructure is under stress because of AI scrapers.
Dumps or no dumps, these AI companies don't care. They feel like they're entitled to taking or stealing what they want.
That's crazy, it makes no sense, it takes as much bandwidth and processing power on the scraper side to process and use the data as it takes to serve it.
They also have an open API that makes scraper entirely unnecessary too.
Here are the relevant quotes from the article you posted
"Scraping has become so prominent that our outgoing bandwidth has increased by 50% in 2024."
"At least 65% of our most expensive requests (the ones that we can’t serve from our caching servers and which are served from the main databases instead) are performed by bots."
"Over the past year, we saw a significant increase in the amount of scraper traffic, and also of related site-stability incidents: Site Reliability Engineers have had to enforce on a case-by-case basis rate limiting or banning of crawlers repeatedly to protect our infrastructure."
And it's wikipedia ! The entire data set is trained INTO the models already, it's not like encyclopedic facts change that often to begin with !
The only thing I imagine is that it is part of a larger ecosystem issue, there the rare case where a dump and API access is so rare, and so untrust worthy that the scrapers are just using scrape for everything, rather than taking the time to save bandwidth by relying on dumps.
Maybe it's consequences from the 2023 API wars, where it was made clear that data repositories would be leveraging their place as pool of knowledge to extract rent from search and AI and places like wikipedia and other wikis and forums are getting hammered as a result of this war.
If the internet wasn't becoming a warzone, there really wouldn't be a need for more than one scraper to scrape a site, even if the site was hostile, like facebook, it only need to be scraped once and then the data could be shared over a torrent swarm efficiently.
The problem isn't that the data is already public.
The problem is that the AI crawlers want to check on it every 5 minutes, even if you try to tell all crawlers that the file is updated daily, or that the file hasn't been updated in a month.
AI crawlers don't care about robots.txt or other helpful hints about what's worth crawling or not, and hints on when it's good time to crawl again.
they won’t have to scrape ?
They don't have to scrape; especially if robots.txt tells them not to.
it’s public data, it’s out there, do you want it public or not ?
Hey, she was wearing a miniskirt, she wanted it, right?
No no no, you don't get to invoke grape imagery to defend copyright.
I know, it hurts when the human shields like wikipedia and the openwrt forums are getting hit, especially when they hand over the goods in dumps. But behind those human shields stand facebook, xitter, amazon, reddit and the rest of big tech garbage and I want tanks to run through them.
So go back to your drawing board and find a solution the tech platform monopolist are made to relinquish our data back to use and the human shields also survive.
My own mother is prisoner in the Zuckerberg data hive and the only way she can get out is brute zucking force into facebook's poop chute.
I think the issue is that the scrapers are fully automatically collecting text, jumping from link to link like a search engine indexer.
If this isn't fertile grounds for a massive class-action lawsuit, I don't know what would be.
whos the defendent, specifically?
No, that's a good point. We all bloody well know there isn't a single provider of LLM's that aren't sucking the entire Internet dry while gleefully ignoring robots.txt and expecting everybody else to pay the bill on their behalf, but the AI providers are getting really good at using other people IPs both to mask their identity and to evade blacklists, which is yet another abusive behavior.
But that's beside your point. So forget the class-action lawsuit in favor of the relevant Ombudsman.
Either way, this cannot go on. Donation-driven open source projects are being driven into the ground by exploding bandwidth and hosting costs, people are being forced to deploy tools like Anubis that eats additional resources - including the resources of every legitimate user. The cumulative damage this is doing is no joke.
how this felt like while reading
And once again a WebApplicationFirewall(WAF) was defeated and it turns out that blocklists and bot detection tools like fail2ban are the way to go...
Who could have seen this coming...