- cross-posted to:
- hackernews
- cross-posted to:
- hackernews
Some thoughts on how useful Anubis really is. Combined with comments I read elsewhere about scrapers starting to solve the challenges, I’m afraid Anubis will be outdated soon and we need something else.
You must log in or register to comment.
This… makes no sense to me. Almost by definition, an AI vendor will have a datacenter full of compute capacity.
Well it doesnt fucking matter what “makes sense to you” because it is working…
Its being deployed by people who had their sites DDoS’d to shit by crawlers and they are very happy with the results so what even is the point of trying to argue here?The current version of Anubis was made as a quick “good enough” solution to an emergency. The article is very enthusiastic about explaining why it shouldn’t work, but completely glosses over the fact that it has worked, at least to an extent where deploying it and maybe inconveniencing some users is preferable to having the entire web server choked out by a flood of indiscriminate scraper requests.
The purpose is to reduce the flood to a manageable level, not to block every single scraper request.
The problem is that the purpose of Anubis was to make crawling more computationally expensive and that crawlers are apparently increasingly prepared to accept that additional cost. One option would be to pile some required cycles on top of what’s currently asked, but it’s a balancing act before it starts to really be an annoyance for the meat popsicle users.
And it was/is for sure the lesser evil compared to what most others did: put the site behind Cloudflare.
I feel people that complain about Anubis have never had their server overheat and shut down on an almost daily basis because of AI scrapers 🤦
Is there a reason other than avoiding infrastructure centralization not to put a web server behind cloudflare?
Unless you have a dirty heatsink, no amount of hammering would make the server overheat
Are you explaining my own server to me? 🙄
What CPU do you have made after 2004 that doesn’t have automatic temperature control ?
I don’t think there is any, unless you somehow managed to disable it ?
Even a raspberry pi without a heatsink won’t overheat to shutdownYou are right, it is actually worse, it usually just overloads the CPU so badly that it starts to throttle and then I can’t even access the server via SSH anymore. But sometimes it also crashes the server so that it reboots, and yes that can happen on modern CPUs as well.
You need to set you http serving process to a priority below the administrative processes (in the place where you are starting it, so assuming linux server that would be your init script or systemd service unit).
Actual crash causing reboot ? Do you have faulty ram maybe ? That’s really not ever supposed to happen from anything happenning in userland. That’s not AI, your stuff might be straight up broken.
Only thing that isn’t broken that could reboot a server is a watchdog timer.
You server shouldn’t crash, reboot or become unreachable from the admin interface even at 100% load and it shouldn’t overheat either, temperatures should never exceed 80C no matter what you do, it’s supposed to be impossible with thermal management, which all processors have had for decades.
Yeah, I’m just wondering what’s going to follow. I just hope everything isn’t going to need to go behind an authwall.
The developer is working on upgrades and better tools. https://xeiaso.net/blog/2025/avoiding-becoming-peg-dependency/
I’ll say the developer is also very responsive. They’re (ambiguous ‘they’, not sure of pronouns) active in a libraries-fighting-bots slack channel I’m on. Libraries have been hit hard by the bots: we have hoards of tasty archives and we don’t have money to throw resources at the problem.
Cool, thanks for posting! Also the reasoning for the image is cool.
Yeah, well-written stuff. I think Anubis will come and go. This beautifully demonstrates and, best of all, quantifies the negligence of Anubis.
It’s very interesting to try to think of what would work, even conceptually. Some sort of purely client-side captcha type of thing perhaps. I keep thinking about it in half-assed ways for minutes at a time.
Maybe something that scrambles the characters of the site according to some random “offset” of some sort, e.g maybe randomly selecting a modulus size and an offset to cycle them, or even just a good ol’ cipher. And the “captcha” consists of a slider that adjusts the offset. You as the viewer know it’s solved when the text becomes something sensical - so there’s no need for the client code to store a readable key that could be used to auto-undo the scrambling. You could maybe even have some values of the slider randomly chosen to produce English text if the scrapers got smart enough to check for legibility (not sure how to hide which slider positions would be these red herring ones though) - which could maybe be enough to trick the scraper into picking up junk text sometimes.
That kind of captcha is trivial to bypass via frequency analysis. Text that looks like language, as opposed to random noise, is very statistically recognisable.
That type of captcha already exists. I don’t know about their specific implementation, but 4chan has it, and it is trivially bypassed by userscripts.
That’s the great thing about Anubis: it’s not client-side. Not entirely anyways. Similar to public key encryption schemes, it exploits the computational complexity of certain functions to solve the challenge. It can’t just say “solved, let me through” because the client has to calculate a number, based on the parameters of the challenge, that fits certain mathematical criteria, and then present it to the server. That’s the “proof of work” component.
A challenge could be something like “find the two prime factors of the semiprime
1522605027922533360535618378132637429718068114961380688657908494580122963258952897654000350692006139”. This number is known as RSA-100, it was first factorized in 1991, which took several days of CPU time, but checking the result is trivial since it’s just integer multiplication. A similar semiprime of 260 decimal digits still hasn’t been factorized to this day. You can’t get around mathematics, no matter how advanced your AI model is.@rtxn I don’t understand how that isn’t client side?
Anything that is client side can be, if not spoofed, then at least delegated to a sub process, and my argument stands
Please, explain to us how you expect to spoof a math problem that you have to provide an answer to the server before proceeding.
You can math all you want on the client, but the server isn’t going to give you shit until you provide the right answer.
@Passerby6497 I really don’t understand the issue here
If there is a challenge to solve, then the server has provided that to the client
There is no way around this, is there?
You’re given the challenge to solve by the server, yes. But just because the challenge is provided to you, that doesn’t mean you can fake your way through it.
You still have to calculate the answer before you can get any farther. You can’t bullshit/spoof your way through the math problem to bypass it, because your correct answer is required to proceed.
There is no way around this, is there?
Unless the server gives you a well-known problem you have the answer to/is easily calculated, or you find a vulnerability in something like Anubis to make it accept a wrong answer, not really. You’re stuck at the interstitial page with a math prompt until you solve it.
Unless I’m misunderstanding your position, I’m not sure what the disconnect is. The original question was about spoofing the challenge client side, but you can’t really spoof the answer to a complicated math problem unless there’s an issue with the server side validation.
@Passerby6497 my stance is that the LLM might recognize that the best way to solve the problem is to run chromium and get the answer from there, then pass it on?
It’s not client-side because validation happens on the server side. The content won’t be displayed until and unless the server receives a valid response, and the challenge is formulated in such a way that calculating a valid answer will always take a long time. It can’t be spoofed because the server will know that the answer is bullshit. In my example, the server will know that the prime factors returned by the client are wrong because their product won’t be equal to the original semiprime. Delegating to a sub-process won’t work either, because what’s the parent process supposed to do? Move on to another piece of content that is also protected by Anubis?
The point is to waste the client’s time and thus reduce the number of requests the server has to handle, not to prevent scraping altogether.
@rtxn validation of what?
This is a typical network thing: client asks for resource, server says here’s a challenge, client responds or doesn’t, has the correct response or not, but has the challenge regardless
THEN (and this is the part you don’t seem to understand) the client process has to waste time solving the challenge, which is, by the way, orders of magnitudes lighter on the server than serving the actual meaningful content, or cancel the request. If a new request is sent during that time, it will still have to waste time solving the challenge. The scraper will get through eventually, but the challenge delays the response and reduces the load on the server because while the scrapers are busy computing, it doesn’t have to serve meaningful content to them.
@rtxn all right, that’s all you had to say initially, rather than try convincing me that the network client was out of the loop: it isn’t, that’s the whole point of Anubis
Yeah has seemed like a bit of a waste of time, once that difficulty gets scaled up and expiration down it’s gonna get annoying to use the web on phones
I had to get my glasses to re-read this comment.
You know why anubis is in place on so many sites, right? You are literally blaming the victims for the absolute bullshit AI is foisting on us all.
I don’t think so. I think he’s blaming the “solution” as being a stop gap at best and painful for end-users at worst. Yes the AI crawlers have caused the issue but I’m not sure this is a great final solution.
As the article discussed, this is essentially “an expensive“ math problem meant to deter AI crawlers but in the end it ain’t really that expensive. It’s more like they put two door handles on a door hoping the bots are too lazy to turn both of them but also severely slowing down all one-handed people. I’m not sure it will ever be feasible to essentially figure out how to have one bot determine if the other end is also a bot without human interaction.
