and a handful of other ASNs on almost all of my properties. It would be one thing if it was simple scraping, but it's often genuine agentic abuse (just spamming for vulnerabilities* or scraping millions of pages with very little valuable or dynamic content). These clouds must get the abuse happening under their own roof under control.
I banned Alibaba's Cloud two years ago and it still hits one of my domains 1 million times per day, despite every route 4xx-ing during that time.
Even Apple has aggressively ramped up its scraping in the past few months, but thankfully they do respect robots.txt.
It's frustrating because some of my real users route through these clouds via their work VPNs. Also, there are a lot of less-than-reputable AI scrapers using global bot-nets via residential IPs to scrape, making it even harder to detect and ban them without affecting real people.
* like leaked .env routes, brute forcing admin panels, and so on
Can you do something with fail2ban? I.e. after a small number of requests resulting in 401 or 404, or a series of requests that's too fast to be reasonably from a human, ban that address for 48 hours.
If a legitimate request later comes from that address, it will probably get through.
But not sure how well iptables handles checking every connection against a list of possibly hundreds of thousands of addresses.
Application-level banning seems very inefficient, these domains should be banned at the network (IP) level. Just drop any incoming packets from these domains/IPs, send absolutely nothing back that even confirms that you exist.
From a domain registry operators point of view Cloudflare are the worst of the bunch by orders of magnitude.
Google DNS are busy but its a hand full of IPs and reasonable. Cloudflare is non-stop never ending traffic. I don't know WTF they are trying to get. I've also seen TCP queries from those same Cloudflare IPs so this is not the result of a DDOS attack on them.
Bots and spiders hang onto old URLs longer than I hold onto a grudge.
We had some SSR that followed up with JavaScript listeners to do interactive updates instead of loading a whole new page. But the team responsible didn’t add nofollow and some bots ignore it anyway. Some significant fraction of our cluster size was answering those requests. But one deployment someone broke the fuck out of those links and it showed up immediately in our 4xx error statistics and we rolled the change back within a week.
Three years later I was still seeing those busted URLs being submitted.
Then the fix for the nofollow got implemented and declared done but the flag never got made permanent. We were pretty peeved when someone learned that in some offhand comment in a conversation between teams.
We were already identifying “good” and “bad” bots at the border and telling them to fuck off, so all of this was coming from our supposed friends. And that’s when I realized that a company owning a search engine and a cloud service provider is a protection racket. Google cloud is selling you servers in order to be talked to by Google bots.
> like leaked .env routes, brute forcing admin panels, and so on
That sort of thing was very common long before agent based abuse scanners. Volume would wax and wane but never went away.
Though other scans for vulnerabilities have shot up in frequency now everyone script kiddie and their flee-bitten mutt can ask an AI to do it for them.
Most big services will just take the junk traffic. They're never going to block whole cloud providers.
Blocking cloud providers is strictly for us dirty self hosters who are sick of our logs becoming unreadable and our visitor metrics useless from all this malicious traffic.
Does the money add up for that to be a deterrent? Are those using them for general VPN access paying enough that they'd care much, and if so is that enough compared to what they'd lose by kicking off the users abusing external sites (they can't all be using just free plans, and anyway tests to kick them could have enough false positives to lose them good paying customers as well as bad users). In any case is the amount of traffic enough for them to care to spend the effort tracking it to identify the bad users?
I’ve had to block similar on clients sites. Their search pages would start to get hammered with obvious agentic search spam. Searching for thing not even related to the site in all combinations of the queries. Some of the user queries contained potentially PII.
However, they could just use residential proxy botnets.
I believe the future is static resources with dynamic apps being only offered to people who can authenticate. And even then I’d recommend grandfathered accounts (eg telegram users that had an account for at least a year).
If there's a know CVE in a version of cPanel ir wordpress for example, those platforms broadcast what version they're on, so the matching cve being known is not illegal. You're just correlating known vulnerable versions without doing any probing. Pen testing however, a bit more involved in pulling info out without permission is illegal.
Can you enforce it? Would you even know who to go after?
Traditionally cybercrimes have been prosecuted by the victim going to their local PD (or the FBI) with the IP address of the offender, and then law enforcement subpoenas the ISP associated with that IP address to figure out the customer associated with that account, and then they have a case for prosecution. It all works as long as cybercriminals are less numerous than law enforcement, fall in the same jurisdiction (or in jurisdictions that have reciprocal extradition treaties), and have a real identity that can be tied to their online activities.
Now, many law enforcement agencies are tied up prosecuting political opponents of their bosses. To the extent that they have bandwidth to investigate and prosecute ordinary crimes, they're going to prioritize the ones that target a.) the government that backs them (of which there are many - you see another headline about the government being hacked every week or so on HN) b.) major corporations that have sophisticated IT systems that can aid prosecution or c.) headlines that will get their boss a PR boost and themselves a promotion. Your personal website does not register among those priorities. You probably can't even prove damages from the intrusion itself. The perpetrator might be a Russian bot that the U.S. doesn't have jurisdiction over, or a Canadian teen that the U.S. previously might've been able to prosecute with reciprocal judicial treaties but now has no leverage over because we've blown up our relationship with Canada. It may be legally a crime, but economically it's not worth prosecution.
Traditionally security researchers had to be very careful because they were doing everything above-board, with their real identity attached to their actions, and they made a very attractive (and PR-worthy) target. Somebody who's port-scanning from Russia through a VPN or residential proxy network is a.) anonymous b.) untraceable c.) out of U.S. jurisdiction and d.) not a headline-worthy target anyway.
A couple years ago I predicted here that a lot of things will break as the U.S. transitions from a high-trust to a low-trust society. This is one of them. Much of our justice system works well when criminals are rare and utterly falls apart when they're ubiquitous.
Last year, I wrote a post about how I use zipbombs to deal with bot traffic [0]. It went mega viral over the web, and I regret it. My server was completely overwhelmed with bot traffic. After months of dealing with it, I had to find a better strategy then keep my mouth shut.
It worked fine... on a small scale But it's one thing to swat flies versus fighting swarm of yellow jackets. Serving a 1MB file a million times a day on a small digitalocean droplet is not feasible. Let's just say it didn't scale. And AI bots usually come with beefier machines, and recover faster after a bomb.
It's the content of the page, the raw html, that is replaced with a zipbomb. Example: https://example.com checks if you are a malicious bot, depending on the result, it will send you gzip html or gzipped bomb. So that can't be swapped or redirected to a CDN. The bots would simply ignore the redirect.
Last year, I read your post and implemented it. I had no way to tell whether there is any benefit to it, so I discontinued it. Serving the data, even though it is compressed and the compression is precomputed, was still a waste of bandwidth. It could hit the browsers of legit user (false positive).
The main problem on my site is that bots hit the CGIT repo, where visiting files triggers syntax highlighting scripts.
I put in a skill-testing-question type system where you get a maximally long-lived access cookie if you answer right (so legit users are bothered as infrequently as possible).
I posted a link to an online game I made in the comments here once, and immediately got all kinds of good and bad traffic.
Some people liked my work, found me on LinkedIn and sent a connection with a positive message. Other people tried every which way to SQL inject or attack the site, posted slurs in every available input, etc.
Someone copied the site and hosted their own version with a couple minor changes. I only found that one because they had left one of the links in, which gave me referring traffic from their copy.
All that from one HN comment, very much pre-AI too, not to mention the standard daily slew of requests probing for what I assume are known vulnerabilities in Wordpress or PHP.
The internet is a floodgate of different actors, many of which are automated, and a link posted somewhere like HN opens the valve especially wide.
I got $10K GCP bill one month because of ClaudeBot. Its likely some users were trying to replicate our software by using Claude code + chrome plugin. If I was spending $10K on traffic, I can only imagine the money cloners must be spending on tokens.
Well, potentially very little! It could be subscription usage, or just really inefficient code authored by an agent. It probably costs LESS in tokens to write a poor quality cloner that breaks upstream things.
It's too late to comment on your older posts, but I'd just like to say: it's nice to see someone making oldschool web browsers work properly on the modern web.
I rolled my own version of Anubis for reasons*, and > 80% of all traffic is bots that fail the challenge, even when white listing legit crawlers.
* I wanted an elixir version so I could run it alongside my existing apps instead of having it sit as a proxy in front of them. This way my existing apps can validate if the caller is already whitelisted and if not redirect to the challenger https://sphinx.8by3.com which sends you back to a site if you pass eg. https://tuhat.net
If you are rolling your own it probably doesn't need to be as clever as Anubis is trying to be. Simple something relatively CPU heavy in JS will do the trick, bonus points if you can make it memory-heavy too. Most bots will give up running something CPU heavy currently (though they will run a lot of JS stuff as much of the web is inaccessible without) and they aren't going to optimise your challenge in WASM or similar because it isn't worth it for the small number of sites it protects (Anubis' challenges on the other hand may well be worth optimising for that way because of their wider use, hence that project is stuck in the cleverness war of attrition and definitely needs to move toward memory-hard problems and something more efficient than JS sooner rather than later).
My main reason for planning to implement something like that myself instead of something off the shelf is two-fold:
1. The points mentioned above
2. I can be a dick and make some subsets do harder work (longer challenges, adding time delays that wouldn't phase a bot) - “hi there North American IP, I know how much you like extra hassle hence the love of effectively paying for tarriffs, here, have a little extra work on me!”
> “hi there North American IP, I know how much you like extra hassle hence the love of effectively paying for tarriffs, here, have a little extra work on me!”
I'm assuming you're outside the US based on having this idea.
You might as well just say "I'm mad at your country for the tariffs, so here's an extra wait." From the fact you feel the urge to spite people in a tariffing country like this, it's clear the tariffs have had some negative effect on you, which goes against your narrative that they're "effectively paid for" by those in the tariffing country who voted for it.
I want mini games! "Score 5 points" style. Although, it would surely fail ADA compliance. I don't see how any of this could be practically effective, respecting ADA.
In server logs you can distinguish the type of LLM's traffic. OAI-SearchBot/GPTBot/ClaudeBot is just indexing but ChatGPT-User/Claude-User etc. may be valuable traffic generated by user's conversation with LLM's. So it's worth choosing which ones you block, instead of blocking all of them.
What's the value for the website maker in these? There's user value to it (rather than just value for Anthropic/OpenAI) but the whole point of these bots is to make sure nobody ever needs to visit your website.
In my case this is the main source of registered users. If a user is engaged in a discussion with chatbot and gets the recommendation from it his engagement is higher. For example, in my system, a user has registered and within one hour used the system 7 times and stayed active while for registrations from Google I can observe zero-activity accounts.
Tons of people click from ChatGPT! It's now the top referrer on most of my websites, and seems to be on path to overcoming Google soon (a ton of traffic from Google has vanished since the AI Overview rollout...).
Strong agree, a good way forward is letting user agent once again truly mean an agent on behalf of the user. And speaking anecdotally, I've seen different traffic "shapes" under the same UA, which leads me to suspect that some actors might be disguising traffic as more mainstream labs.
I have tried a ton of stuff (Cloudflare, Anubis, fail2ban...) to combat the bots lately, but they all have their downsides, and users complained too much, so I just gave up. You simply can't win this arms race.
So I have resorted to performance improvements on my server and serving every single request. Things are way smoother now.
It's unfortunate that this also just leads to port scanning and huge accidental bills. This happened in the past to more than one acquaintnace of mine; go and market their vibe-coded/OS project on HN + twitter and then suddenly the vercel bill goes through the roof with bots scraping the website to bits, even with bot protection from CF. Really hard to get away from it these days, but hopefully over the next few years some new startup will raise $50M for "killing cloudflare" and will actually follow through!
I posted a link to HN a few days ago that got literally zero upvotes and zero comments (Applying NASA's satellite-imaging trick to ancient museum objects https://news.ycombinator.com/item?id=50023128) and it was hit with insane bot traffic. HN was the only place I had posted it.
> I won't name names, but I looked them up, and they describe themselves as a service that listens for changes to a website, and then summarizes them into an AI slop news letter. And by 'listens to changes' I mean query our website many many times per-minute. Blocked.
That sort of behavior would justify a name and shame imo
Inconsiderate shitty services don’t deserve the politeness of protecting their name
I'm trying to understand the reasoning behind the entity that was sending presumably hundreds or thousands of requests per minute to the same page. It's some kind of AI news site, so it makes sense they want the freshest content, but it sounds like they were doing many requests per second, so that can't be the reason?
One option to reduce it, though this is work to setup initially:
1. Use a catch-all subdomain
2. Reject email to addresses where a simple checksum does not match (valid addresses being something like asdfghij.klmn@catchall.domain.tld where the klmn is taken from hash of asdfghij and some secret, anything not matching that pattern gets rejected) [this means running your own mail server so you can do that on incoming mail, running a scan to pick out and remove errant messages regularly, or having a mail client where you can implement this as a plugin]
3. On your contact detail page, don't give out a straight email address but have the the page run some Anubis style JS before asking the server to respond with an address that works
4. Once an address is obviously being used in a manner you don't care to respond to, add it to an explicit block list.
For extra protection from addresses being passed around (intentionally or via hack & other leaks), ask for their email address first and use that as part of the hash, then you can reject mail to your catch-all that doesn't come from the original other party.
Whether all this is worth the effort of implementing, is of course up to you. It probably isn't… I've been considering something like it, but mainly as a fun little project rather than something that would be massively useful.
You shouldnt defend bad actors and abusers. Hiding them is tantamount to them attacking others indiscriminately. Cause youre not the first victim, and definitely not the last.
You see something, you say something, or you do something.
Certificate transparency logs get scraped. However, they cannot scrape wildcards. Get a domain.example + *.domain.example cert and host your blog on blog.domain.example and you can keep the bots from finding you, at least for a little while.
No, it wouldn't. Just yesterday I added a domain.example + *.domain.example certificate, and today some bot comes to scan for subdomains: api, app, backup, beta, cdn, ci, mail etc. Blog isn't in the list, but just for this bot.
Turnstile is no longer effective. Anyone with a budget uses a real headless chrome instance and gets around it. You only need the spendy chrome instances to get the initial cookie, then you pass that to your workers and they scrape.
It’s really starting to suck. The worst part is these are coming from consumer vpn providers that we can’t just block (lots of real users paying real money on our service are vpn users) or they come from big residential isps. The whole residential relay thing is a hard adversary, since it’s spread across so many real users paying ips that you can’t whackamole it. I’d say 1% of residential ips are participating in a botnet (there’s prob some real numbers out there if I cared to look).
This is just going to push everyone harder into walled gardens. Web input biggest inbound but our content is our mote, we are going to have to turn off non logged in content viewing, which will have real business impacts.
I run a few private MMO servers, we have a website with a support forum, in order to post, you need to register, log in, fill in account data, then create a thread, each of these pages is protected by Turnstile as well as a general bot check for the website, we frequently get automated spam from RMT (real-money trade) services for the OG game. Turnstile is maybe effective against simple bots that just try to call curl/requests/whatever but if you have bots that use a browser driver it's completely useless.
I did this for alyph.ai, but was recently flooded with so many browser-using, Turnstile-defeating bots from Taiwan that I had to completely turn off sign ups. It's not fun anymore.
Not even third party login worked (Google, Linkedin, ...) to curb this.
They were logging in, scraping the client API for vulnerabilities/possibilities and then spammed erotic dog poems to Claude Fable until the $5 welcome credit got drained.
I've been posting about this recently but I continue to be bad at either timing my submissions or titling them. I struggle with YouTube thumbnails too so I think it might be me. But anyway:
I experienced the same thing after submitting articles from my small photography website. I did debate using some of the off-the-shelf solutions like Cloudflare, Anubis, or iocaine, but I ended up building my own L7 proxy that utterly fucks the bots up: https://robertmay.photography/journal/bwbach-my-guardian-gob.... I will make it generally available at some point once I get it optimised how I'd like, but currently I'm able to trap about 3000 concurrent crawlers in my tarpit of slow requests on two small Hetzner VMs. And that 3k is really just the limits of the networks hitting my servers right now, I think it'll go significantly higher than that.
I haven't really seen anyone else combat them this way, possibly because many of the standard proxies and servers are set up for fast requests, not serving them as slowly as possible. Deliberately serving page loads for as long as the requester will accept packets turns out to be a really funny way to stop the botnets - when I just rejected the requests they in turn hammered my server even harder. By tanking the request volume with a ridiculous BEAM-based proxy you can reduce their request speed to a crawl and I'm fairly certain a couple of them have blacklisted my server now.
I like the spirit of this approach; make them suffer! But I'm surprised the bots wouldn't have reasonable timeouts to counteract. How long do they typically wait on these requests?
EDIT: sorry, that was a dumb question before reading your post. You actually learn the timeout they tolerate. Nice!
I should probably put together some actual stats at some point, but from what I see in the live logs you'd be shocked - some of them will happily sit there for over 5 minutes being drip-fed packets!
Looking at you, Meta.
I do still get some requests going up to 15 minutes which is quite ridiculous, though those don't seem to be super common (maybe they're learning).
I think the shortest timeouts I've seen are between 2-5 seconds, but they're surprisingly infrequent and actually that's still a massive inconvenience for them. I did at one point just drop any request I couldn't keep open for more than 5 seconds but it actually caused more server load. I'm essentially inverting the server load by making slow requests more efficient, which is counter to everything I ever learnt about performance for web serving :D
I think the average I see is probably somewhere between 15-30 seconds. I'm guessing many crawlers don't want to go too low because they'll be trying to fetch large files like images, or perhaps they know they'll be slowing servers down and want to ensure they get the response. Most likely, the majority of crawler authors aren't expecting some arsehole to exploit this and they've just left the setting as the default.
What confused me most about this is that the entire site is, apparently, a single page (there's no sitemap that I can use to verify this; there IS an RSS file which appears to confirm it) containing all posts. Links at the top (headed "Table of Contents") are all in-page links.
At first, I assumed this was a single blog post, especially with a "Table of Contents" at the top. I guess it's following some old-school blogging conventions, but AFAICR, even those sites retained individual pages for separate posts.
You can also link between Org documents and org-publish will export the documents with working web links, if you're lucky. That's how https://entropicthoughts.com works.
If you have an SSL certificate so such a domain can serve HTTPS traffic if needed, which it wouldn't surprise me if some parking services generate for you these days via an ACME provider, then your domain will be in the certificate transparency logs so can be found that way.
One possible way around this is to use a wildcard and not serve anything on the base domain. Unless the right subdomain is listed somewhere, they'll not get any scans. Maybe useful for projects that you have online for just yourself and a small closed circle.
I wonder if this is why I see so many YouTube links posted under ‘new’ that never make it to the FP. Is this a way to get bot scraped and boost views/interactions?
Possibly, possibly not. I think some people are used to passing around news/information as short-form videos these days but HN (mostly) isn't into that kind of content.
>So I took a look at some data. I added top-level logging for every endpoint (how many times was a request made to each url path), and I found that a majority of the requests did not even make an attempt to fetch news articles! For reference, my website is built as a single-page-application, and what this means is that a majority of the requests did not even run the javascript necessary to render the page (!).
So you got hammered by curl requests! But isn't that part of your delivery static anyway? Were you regenerating the same HTML over and over?
>many, many requests per minute?
How many was it?
Also interesting that it was a single "user". (user-agent?) Did it come from a different IP every time?
> So you got hammered by curl requests! But isn't that part of your delivery static anyway? Were you regenerating the same HTML over and over?
i dont think this was the cause of the slowdown, it was just sort of the thing that led to the discovery that this was mostly bots.
in general, because the purpose of the website is to allow users to generate news feeds based on custom criteria, it does do a bit of meaningful work on each request (ones that do follow through and load the full page).
> How many was it?
> Also interesting that it was a single "user". (user-agent?) Did it come from a different IP every time?
I'm intentionally leaving some of the details vague here, both because I don't want to be too harsh on the original author (it may have just been accidental/a buggy program/etc.), and because I don't want the traffic generators to know how im flagging them.
I recently posted a link that got 2 boosts and 6 faves, but it was requested by about 300 fedi servers so they could generate a preview.
Which is fine! Because even though my photo gallery runs on a 15 year old laptop in the basement, I wrote the gallery software in a fast language (Rust) and so a trickle of requests maxing at 4 per second is not actually a problem.
People complaining about fedi hammering their servers are probably using PHP or Python.
Ok, on the one hand to have one bot hitting your website to populate some kind of slop newsletter is pretty annoying.
On the other hand, “multiple times per minute” doesn’t seem that high? Even if it was 60x per min that’s fully within the limits of what a 2015 laptop in the closet can handle.
Glad you're liking unslop.news :) If you ever find any bugs or want anything added, it's open source so feel free to open a ticket https://github.com/leiDnedyA/hn-without-ai
I _also_ posted to HN and made it to #1[0]; the effect this had on my CPU/network traffic on a 5/EUR month VPS over at Hetzner was in the single digit percentages.
As per the wiki page for the C10k problem[1] where it says "_was_ [emphasis mine] the problem of optimizing computer networking stacks to handle a large number of clients at the same time". This was solved over a decade ago...
The c10k problem is a different problem. It relates to simultaneous open TCP connections at a server. It would take a very very large number of crawler bots to force open 10k connections, I suspect.
There is so much spam, abuse, and aggressive scraping that I've had to block:
and a handful of other ASNs on almost all of my properties. It would be one thing if it was simple scraping, but it's often genuine agentic abuse (just spamming for vulnerabilities* or scraping millions of pages with very little valuable or dynamic content). These clouds must get the abuse happening under their own roof under control.I banned Alibaba's Cloud two years ago and it still hits one of my domains 1 million times per day, despite every route 4xx-ing during that time.
Even Apple has aggressively ramped up its scraping in the past few months, but thankfully they do respect robots.txt.
It's frustrating because some of my real users route through these clouds via their work VPNs. Also, there are a lot of less-than-reputable AI scrapers using global bot-nets via residential IPs to scrape, making it even harder to detect and ban them without affecting real people.
* like leaked .env routes, brute forcing admin panels, and so on
> I banned Alibaba's Cloud two years ago and it still hits one of my domains 1 million times per day, despite every route 4xx-ing during that time.
You may as well silently drop the packets if they don't adapt.
Having the same issue with Bing bot. No idea why it's so eager to not behave.
Microsoft has never understood consent
Can you do something with fail2ban? I.e. after a small number of requests resulting in 401 or 404, or a series of requests that's too fast to be reasonably from a human, ban that address for 48 hours.
If a legitimate request later comes from that address, it will probably get through.
But not sure how well iptables handles checking every connection against a list of possibly hundreds of thousands of addresses.
Fail2ban doesn’t work well against residential bot nets. A single ip generally only makes a couple of requests before rotating.
ipsets can handle vast numbers of addresses. I'm sure fail2ban uses them these days.
Application-level banning seems very inefficient, these domains should be banned at the network (IP) level. Just drop any incoming packets from these domains/IPs, send absolutely nothing back that even confirms that you exist.
They are banned at the edge in Cloudflare (now), they never hit my application.
From a domain registry operators point of view Cloudflare are the worst of the bunch by orders of magnitude.
Google DNS are busy but its a hand full of IPs and reasonable. Cloudflare is non-stop never ending traffic. I don't know WTF they are trying to get. I've also seen TCP queries from those same Cloudflare IPs so this is not the result of a DDOS attack on them.
> Cloudflare is non-stop never ending traffic. I don't know WTF they are trying to get. I've also seen TCP queries from those same Cloudflare IPs
1. Breach router or desktop that hasn't been patched in ages via mass script
2. Establish C&C using cloudflare tunnels
3. Sell these hacked systems as residential proxy endpoints to everybody from state actors to teenagers
4. Some of them will be data scraping services
You are chasing your tail trying to ban by ASN or IP. It may work for another couple of months, at best.
Makes one want to build a parallel web based on web of trust.
Cloudflare do a great job, I noticed moving from another VM.
at controlling internet sure
Bots and spiders hang onto old URLs longer than I hold onto a grudge.
We had some SSR that followed up with JavaScript listeners to do interactive updates instead of loading a whole new page. But the team responsible didn’t add nofollow and some bots ignore it anyway. Some significant fraction of our cluster size was answering those requests. But one deployment someone broke the fuck out of those links and it showed up immediately in our 4xx error statistics and we rolled the change back within a week.
Three years later I was still seeing those busted URLs being submitted.
Then the fix for the nofollow got implemented and declared done but the flag never got made permanent. We were pretty peeved when someone learned that in some offhand comment in a conversation between teams.
We were already identifying “good” and “bad” bots at the border and telling them to fuck off, so all of this was coming from our supposed friends. And that’s when I realized that a company owning a search engine and a cloud service provider is a protection racket. Google cloud is selling you servers in order to be talked to by Google bots.
> like leaked .env routes, brute forcing admin panels, and so on
That sort of thing was very common long before agent based abuse scanners. Volume would wax and wane but never went away.
Though other scans for vulnerabilities have shot up in frequency now everyone script kiddie and their flee-bitten mutt can ask an AI to do it for them.
> These clouds must get the abuse happening under their own roof under control.
Similar to ad click fraud and scam ads, the cloud provider gets paid, why would they turn down the money?
So people accessing the internet through their cloud don't get automatically ip-banned by most services, perhaps?
Most big services will just take the junk traffic. They're never going to block whole cloud providers.
Blocking cloud providers is strictly for us dirty self hosters who are sick of our logs becoming unreadable and our visitor metrics useless from all this malicious traffic.
Does the money add up for that to be a deterrent? Are those using them for general VPN access paying enough that they'd care much, and if so is that enough compared to what they'd lose by kicking off the users abusing external sites (they can't all be using just free plans, and anyway tests to kick them could have enough false positives to lose them good paying customers as well as bad users). In any case is the amount of traffic enough for them to care to spend the effort tracking it to identify the bad users?
If advertising and tracking was less egregious, I wouldn't feel the need to use an ad blocker. But, here we are.
Sorry, what's the connection with the topic at hand?
Meta went absolutely haywire on one of my sites a couple months ago.
I’ve had to block similar on clients sites. Their search pages would start to get hammered with obvious agentic search spam. Searching for thing not even related to the site in all combinations of the queries. Some of the user queries contained potentially PII.
Is there a handy list of all of this that users contribute to?
This is a very smart thing to do! Thanks.
Why allow random requests coming from the clouds?
However, they could just use residential proxy botnets.
I believe the future is static resources with dynamic apps being only offered to people who can authenticate. And even then I’d recommend grandfathered accounts (eg telegram users that had an account for at least a year).
> However, they could just use residential proxy botnets.
Agreed
> I believe the future is static resources with dynamic apps being only offered to people who can authenticate.
I don't know about static resources, but I do think that identity will come into play.
AWS is terrible. We had another tenant on AWS attacking us on AWS. AWS support took nearly a week of slowly bouncing around a case to do anything.
This is a very smart thing to do! Thanks.
Why allow random requests coming from the clouds?
However, they could just use residential proxy botnets.
Isn’t scanning for vulnerabilities illegal? Traditionally security researchers had to be very careful about such things
If there's a know CVE in a version of cPanel ir wordpress for example, those platforms broadcast what version they're on, so the matching cve being known is not illegal. You're just correlating known vulnerable versions without doing any probing. Pen testing however, a bit more involved in pulling info out without permission is illegal.
Can you enforce it? Would you even know who to go after?
Traditionally cybercrimes have been prosecuted by the victim going to their local PD (or the FBI) with the IP address of the offender, and then law enforcement subpoenas the ISP associated with that IP address to figure out the customer associated with that account, and then they have a case for prosecution. It all works as long as cybercriminals are less numerous than law enforcement, fall in the same jurisdiction (or in jurisdictions that have reciprocal extradition treaties), and have a real identity that can be tied to their online activities.
Now, many law enforcement agencies are tied up prosecuting political opponents of their bosses. To the extent that they have bandwidth to investigate and prosecute ordinary crimes, they're going to prioritize the ones that target a.) the government that backs them (of which there are many - you see another headline about the government being hacked every week or so on HN) b.) major corporations that have sophisticated IT systems that can aid prosecution or c.) headlines that will get their boss a PR boost and themselves a promotion. Your personal website does not register among those priorities. You probably can't even prove damages from the intrusion itself. The perpetrator might be a Russian bot that the U.S. doesn't have jurisdiction over, or a Canadian teen that the U.S. previously might've been able to prosecute with reciprocal judicial treaties but now has no leverage over because we've blown up our relationship with Canada. It may be legally a crime, but economically it's not worth prosecution.
Traditionally security researchers had to be very careful because they were doing everything above-board, with their real identity attached to their actions, and they made a very attractive (and PR-worthy) target. Somebody who's port-scanning from Russia through a VPN or residential proxy network is a.) anonymous b.) untraceable c.) out of U.S. jurisdiction and d.) not a headline-worthy target anyway.
A couple years ago I predicted here that a lot of things will break as the U.S. transitions from a high-trust to a low-trust society. This is one of them. Much of our justice system works well when criminals are rare and utterly falls apart when they're ubiquitous.
Last year, I wrote a post about how I use zipbombs to deal with bot traffic [0]. It went mega viral over the web, and I regret it. My server was completely overwhelmed with bot traffic. After months of dealing with it, I had to find a better strategy then keep my mouth shut.
[0]: https://news.ycombinator.com/item?id=43826798
So it ended up not having a noticeable impact on the bots?
It worked fine... on a small scale But it's one thing to swat flies versus fighting swarm of yellow jackets. Serving a 1MB file a million times a day on a small digitalocean droplet is not feasible. Let's just say it didn't scale. And AI bots usually come with beefier machines, and recover faster after a bomb.
Why did you serve it from DigitalOcean?
Why not a CDN?
It's the content of the page, the raw html, that is replaced with a zipbomb. Example: https://example.com checks if you are a malicious bot, depending on the result, it will send you gzip html or gzipped bomb. So that can't be swapped or redirected to a CDN. The bots would simply ignore the redirect.
That sounds really annoying. Thanks for explaining! I'm looking to launch one of my new little products and my last one got hammered by bots.
Thankfully they were not really malicious like some of the other folks have experienced, but it was still frustrating.
Last year, I read your post and implemented it. I had no way to tell whether there is any benefit to it, so I discontinued it. Serving the data, even though it is compressed and the compression is precomputed, was still a waste of bandwidth. It could hit the browsers of legit user (false positive).
The main problem on my site is that bots hit the CGIT repo, where visiting files triggers syntax highlighting scripts.
I put in a skill-testing-question type system where you get a maximally long-lived access cookie if you answer right (so legit users are bothered as infrequently as possible).
Can that technique be applied to packet compression?
What is that improved strategy? :pray:
Not telling you this time, Mr. Bot
I really recommend blocking all traffic from anything on https://github.com/borestad/blocklist-abuseipdb
I posted a link to an online game I made in the comments here once, and immediately got all kinds of good and bad traffic.
Some people liked my work, found me on LinkedIn and sent a connection with a positive message. Other people tried every which way to SQL inject or attack the site, posted slurs in every available input, etc.
Someone copied the site and hosted their own version with a couple minor changes. I only found that one because they had left one of the links in, which gave me referring traffic from their copy.
All that from one HN comment, very much pre-AI too, not to mention the standard daily slew of requests probing for what I assume are known vulnerabilities in Wordpress or PHP.
The internet is a floodgate of different actors, many of which are automated, and a link posted somewhere like HN opens the valve especially wide.
I got $10K GCP bill one month because of ClaudeBot. Its likely some users were trying to replicate our software by using Claude code + chrome plugin. If I was spending $10K on traffic, I can only imagine the money cloners must be spending on tokens.
Well, potentially very little! It could be subscription usage, or just really inefficient code authored by an agent. It probably costs LESS in tokens to write a poor quality cloner that breaks upstream things.
I am trying to do the same thing, but on purpose for https://notforhumans.pages.dev/
It's too late to comment on your older posts, but I'd just like to say: it's nice to see someone making oldschool web browsers work properly on the modern web.
https://www.sammystraus.com/#using-llms-to-make-the-emacs-we...
:) ive been dreaming of a usable emacs browser for so long
I rolled my own version of Anubis for reasons*, and > 80% of all traffic is bots that fail the challenge, even when white listing legit crawlers.
* I wanted an elixir version so I could run it alongside my existing apps instead of having it sit as a proxy in front of them. This way my existing apps can validate if the caller is already whitelisted and if not redirect to the challenger https://sphinx.8by3.com which sends you back to a site if you pass eg. https://tuhat.net
If you are rolling your own it probably doesn't need to be as clever as Anubis is trying to be. Simple something relatively CPU heavy in JS will do the trick, bonus points if you can make it memory-heavy too. Most bots will give up running something CPU heavy currently (though they will run a lot of JS stuff as much of the web is inaccessible without) and they aren't going to optimise your challenge in WASM or similar because it isn't worth it for the small number of sites it protects (Anubis' challenges on the other hand may well be worth optimising for that way because of their wider use, hence that project is stuck in the cleverness war of attrition and definitely needs to move toward memory-hard problems and something more efficient than JS sooner rather than later).
My main reason for planning to implement something like that myself instead of something off the shelf is two-fold:
1. The points mentioned above
2. I can be a dick and make some subsets do harder work (longer challenges, adding time delays that wouldn't phase a bot) - “hi there North American IP, I know how much you like extra hassle hence the love of effectively paying for tarriffs, here, have a little extra work on me!”
> “hi there North American IP, I know how much you like extra hassle hence the love of effectively paying for tarriffs, here, have a little extra work on me!”
I'm assuming you're outside the US based on having this idea.
You might as well just say "I'm mad at your country for the tariffs, so here's an extra wait." From the fact you feel the urge to spite people in a tariffing country like this, it's clear the tariffs have had some negative effect on you, which goes against your narrative that they're "effectively paid for" by those in the tariffing country who voted for it.
I want mini games! "Score 5 points" style. Although, it would surely fail ADA compliance. I don't see how any of this could be practically effective, respecting ADA.
In server logs you can distinguish the type of LLM's traffic. OAI-SearchBot/GPTBot/ClaudeBot is just indexing but ChatGPT-User/Claude-User etc. may be valuable traffic generated by user's conversation with LLM's. So it's worth choosing which ones you block, instead of blocking all of them.
> valuable traffic
What's the value for the website maker in these? There's user value to it (rather than just value for Anthropic/OpenAI) but the whole point of these bots is to make sure nobody ever needs to visit your website.
In my case this is the main source of registered users. If a user is engaged in a discussion with chatbot and gets the recommendation from it his engagement is higher. For example, in my system, a user has registered and within one hour used the system 7 times and stayed active while for registrations from Google I can observe zero-activity accounts.
Tons of people click from ChatGPT! It's now the top referrer on most of my websites, and seems to be on path to overcoming Google soon (a ton of traffic from Google has vanished since the AI Overview rollout...).
Strong agree, a good way forward is letting user agent once again truly mean an agent on behalf of the user. And speaking anecdotally, I've seen different traffic "shapes" under the same UA, which leads me to suspect that some actors might be disguising traffic as more mainstream labs.
I have tried a ton of stuff (Cloudflare, Anubis, fail2ban...) to combat the bots lately, but they all have their downsides, and users complained too much, so I just gave up. You simply can't win this arms race.
So I have resorted to performance improvements on my server and serving every single request. Things are way smoother now.
It's unfortunate that this also just leads to port scanning and huge accidental bills. This happened in the past to more than one acquaintnace of mine; go and market their vibe-coded/OS project on HN + twitter and then suddenly the vercel bill goes through the roof with bots scraping the website to bits, even with bot protection from CF. Really hard to get away from it these days, but hopefully over the next few years some new startup will raise $50M for "killing cloudflare" and will actually follow through!
I used to roll my own (respectful) scrapers. AI and vibecoders have ruined it for me. Everyone has super sensitive rate limiting and not checking now.
I posted a link to HN a few days ago that got literally zero upvotes and zero comments (Applying NASA's satellite-imaging trick to ancient museum objects https://news.ycombinator.com/item?id=50023128) and it was hit with insane bot traffic. HN was the only place I had posted it.
> I won't name names, but I looked them up, and they describe themselves as a service that listens for changes to a website, and then summarizes them into an AI slop news letter. And by 'listens to changes' I mean query our website many many times per-minute. Blocked.
That sort of behavior would justify a name and shame imo
Inconsiderate shitty services don’t deserve the politeness of protecting their name
I wonder if you could just blast them with ever changing content to burn their summarization tokens on BS.
Though they probably use the cheapest possible models they can so the damage from their dumb business model would be limited
I'm trying to understand the reasoning behind the entity that was sending presumably hundreds or thousands of requests per minute to the same page. It's some kind of AI news site, so it makes sense they want the freshest content, but it sounds like they were doing many requests per second, so that can't be the reason?
> I'm trying to understand the reasoning behind the entity that was sending presumably hundreds or thousands of requests per minute to the same page.
The most likely answer is they don't give a crap because it's not their money and no one is going to do anything about it.
Nice. How do I get a massive swarm of nodes on someone else's money?
Oh is that why I keep getting those ai emails about the project I posted earlier?
Yeah it's awful they keep coming months after you post anything as a Show HN now, definitely discourages me from doing it
One option to reduce it, though this is work to setup initially:
1. Use a catch-all subdomain
2. Reject email to addresses where a simple checksum does not match (valid addresses being something like asdfghij.klmn@catchall.domain.tld where the klmn is taken from hash of asdfghij and some secret, anything not matching that pattern gets rejected) [this means running your own mail server so you can do that on incoming mail, running a scan to pick out and remove errant messages regularly, or having a mail client where you can implement this as a plugin]
3. On your contact detail page, don't give out a straight email address but have the the page run some Anubis style JS before asking the server to respond with an address that works
4. Once an address is obviously being used in a manner you don't care to respond to, add it to an explicit block list.
For extra protection from addresses being passed around (intentionally or via hack & other leaks), ask for their email address first and use that as part of the hash, then you can reject mail to your catch-all that doesn't come from the original other party.
Whether all this is worth the effort of implementing, is of course up to you. It probably isn't… I've been considering something like it, but mainly as a fun little project rather than something that would be massively useful.
> my website is built as a single-page-application
There you go. Ditch it. Make you pages cachable as much as you can.
> one specific user. I won't name names
Who is?!
haha im not naming names
Is it SpokenFeeds?
You shouldnt defend bad actors and abusers. Hiding them is tantamount to them attacking others indiscriminately. Cause youre not the first victim, and definitely not the last.
You see something, you say something, or you do something.
How else will the author play the victim for some clicks then?
Does the country end with an "a" with 5 - 6 characters long?
Oh, as soon as you add tls to your subdomain, the bots come. It’s instant.
Certificate transparency logs get scraped. However, they cannot scrape wildcards. Get a domain.example + *.domain.example cert and host your blog on blog.domain.example and you can keep the bots from finding you, at least for a little while.
No, it wouldn't. Just yesterday I added a domain.example + *.domain.example certificate, and today some bot comes to scan for subdomains: api, app, backup, beta, cdn, ci, mail etc. Blog isn't in the list, but just for this bot.
Have you tried dealing with bots using CF Turnstile? Ref: https://www.cloudflare.com/products/turnstile/
Turnstile is no longer effective. Anyone with a budget uses a real headless chrome instance and gets around it. You only need the spendy chrome instances to get the initial cookie, then you pass that to your workers and they scrape.
It’s really starting to suck. The worst part is these are coming from consumer vpn providers that we can’t just block (lots of real users paying real money on our service are vpn users) or they come from big residential isps. The whole residential relay thing is a hard adversary, since it’s spread across so many real users paying ips that you can’t whackamole it. I’d say 1% of residential ips are participating in a botnet (there’s prob some real numbers out there if I cared to look).
This is just going to push everyone harder into walled gardens. Web input biggest inbound but our content is our mote, we are going to have to turn off non logged in content viewing, which will have real business impacts.
I run a few private MMO servers, we have a website with a support forum, in order to post, you need to register, log in, fill in account data, then create a thread, each of these pages is protected by Turnstile as well as a general bot check for the website, we frequently get automated spam from RMT (real-money trade) services for the OG game. Turnstile is maybe effective against simple bots that just try to call curl/requests/whatever but if you have bots that use a browser driver it's completely useless.
I did this for alyph.ai, but was recently flooded with so many browser-using, Turnstile-defeating bots from Taiwan that I had to completely turn off sign ups. It's not fun anymore.
Not even third party login worked (Google, Linkedin, ...) to curb this.
They were logging in, scraping the client API for vulnerabilities/possibilities and then spammed erotic dog poems to Claude Fable until the $5 welcome credit got drained.
Fun New World.
remove free signups
$5 minimum to sign up
not yet, but I think something like this would be a next step
if you want the site to be fast, don't render it with javascript and take a look at the google lighthouse check for things to improve:
https://pagespeed.web.dev/analysis/https-sprinklz-io/bmc4mv1...
I've been posting about this recently but I continue to be bad at either timing my submissions or titling them. I struggle with YouTube thumbnails too so I think it might be me. But anyway:
I experienced the same thing after submitting articles from my small photography website. I did debate using some of the off-the-shelf solutions like Cloudflare, Anubis, or iocaine, but I ended up building my own L7 proxy that utterly fucks the bots up: https://robertmay.photography/journal/bwbach-my-guardian-gob.... I will make it generally available at some point once I get it optimised how I'd like, but currently I'm able to trap about 3000 concurrent crawlers in my tarpit of slow requests on two small Hetzner VMs. And that 3k is really just the limits of the networks hitting my servers right now, I think it'll go significantly higher than that.
I haven't really seen anyone else combat them this way, possibly because many of the standard proxies and servers are set up for fast requests, not serving them as slowly as possible. Deliberately serving page loads for as long as the requester will accept packets turns out to be a really funny way to stop the botnets - when I just rejected the requests they in turn hammered my server even harder. By tanking the request volume with a ridiculous BEAM-based proxy you can reduce their request speed to a crawl and I'm fairly certain a couple of them have blacklisted my server now.
I'm running some tests currently
https://deep.43z.one (how deep bots will crawl)
https://stalling.43z.one (how long they'll keep the connection open)
I like the spirit of this approach; make them suffer! But I'm surprised the bots wouldn't have reasonable timeouts to counteract. How long do they typically wait on these requests?
EDIT: sorry, that was a dumb question before reading your post. You actually learn the timeout they tolerate. Nice!
I should probably put together some actual stats at some point, but from what I see in the live logs you'd be shocked - some of them will happily sit there for over 5 minutes being drip-fed packets!
Looking at you, Meta.
I do still get some requests going up to 15 minutes which is quite ridiculous, though those don't seem to be super common (maybe they're learning).
I think the shortest timeouts I've seen are between 2-5 seconds, but they're surprisingly infrequent and actually that's still a massive inconvenience for them. I did at one point just drop any request I couldn't keep open for more than 5 seconds but it actually caused more server load. I'm essentially inverting the server load by making slow requests more efficient, which is counter to everything I ever learnt about performance for web serving :D
I think the average I see is probably somewhere between 15-30 seconds. I'm guessing many crawlers don't want to go too low because they'll be trying to fetch large files like images, or perhaps they know they'll be slowing servers down and want to ensure they get the response. Most likely, the majority of crawler authors aren't expecting some arsehole to exploit this and they've just left the setting as the default.
> 3000 concurrent crawlers
I gave up on tarpits when it got to 130k "waiting" connections.
What confused me most about this is that the entire site is, apparently, a single page (there's no sitemap that I can use to verify this; there IS an RSS file which appears to confirm it) containing all posts. Links at the top (headed "Table of Contents") are all in-page links.
At first, I assumed this was a single blog post, especially with a "Table of Contents" at the top. I guess it's following some old-school blogging conventions, but AFAICR, even those sites retained individual pages for separate posts.
it's org-html-export-to-html babyyy lmao
I've wanted to work with org mode for so long, but I can't simply bring myself to use emacs. Nothing personal. Org mode is awesome, tho.
I think I could improve the format a bit with something like Jekyll, but I haven't had a good chunk of time to sit down and work on it
You can also link between Org documents and org-publish will export the documents with working web links, if you're lucky. That's how https://entropicthoughts.com works.
I get llm-like traffic on domains with zero web content (never was, probably never will be any).
Just merely existing will get you scanned these days.
If you have an SSL certificate so such a domain can serve HTTPS traffic if needed, which it wouldn't surprise me if some parking services generate for you these days via an ACME provider, then your domain will be in the certificate transparency logs so can be found that way.
One possible way around this is to use a wildcard and not serve anything on the base domain. Unless the right subdomain is listed somewhere, they'll not get any scans. Maybe useful for projects that you have online for just yourself and a small closed circle.
I wonder if this is why I see so many YouTube links posted under ‘new’ that never make it to the FP. Is this a way to get bot scraped and boost views/interactions?
Possibly, possibly not. I think some people are used to passing around news/information as short-form videos these days but HN (mostly) isn't into that kind of content.
Interesting, thanks for sharing.
>So I took a look at some data. I added top-level logging for every endpoint (how many times was a request made to each url path), and I found that a majority of the requests did not even make an attempt to fetch news articles! For reference, my website is built as a single-page-application, and what this means is that a majority of the requests did not even run the javascript necessary to render the page (!).
So you got hammered by curl requests! But isn't that part of your delivery static anyway? Were you regenerating the same HTML over and over?
>many, many requests per minute?
How many was it?
Also interesting that it was a single "user". (user-agent?) Did it come from a different IP every time?
But hey. Silver lining: free load testing! ;)
> So you got hammered by curl requests! But isn't that part of your delivery static anyway? Were you regenerating the same HTML over and over?
i dont think this was the cause of the slowdown, it was just sort of the thing that led to the discovery that this was mostly bots.
in general, because the purpose of the website is to allow users to generate news feeds based on custom criteria, it does do a bit of meaningful work on each request (ones that do follow through and load the full page).
> How many was it? > Also interesting that it was a single "user". (user-agent?) Did it come from a different IP every time?
I'm intentionally leaving some of the details vague here, both because I don't want to be too harsh on the original author (it may have just been accidental/a buggy program/etc.), and because I don't want the traffic generators to know how im flagging them.
I ran into similar problems. My analytics became so useless that I threw them out.
That is primarily why websites only TLS-negotiates for server-side-only ChaCha-Poly.
How could the specifics of TLS help deter/block bots?
love the simple css of this. very tasteful and beautiful, even if its just simple.
Same with me, every-time I got 1st page I got a lot of robots
Wait until you answer in "Ask HN: Who wants to be hired?" and start receiving hiring phishing emails.
Try posting something viral on Mastodon. The logs clearly show that it isn't the most efficient social network.
Can you elaborate more, please? I'm curious.
I recently posted a link that got 2 boosts and 6 faves, but it was requested by about 300 fedi servers so they could generate a preview.
Which is fine! Because even though my photo gallery runs on a 15 year old laptop in the basement, I wrote the gallery software in a fast language (Rust) and so a trickle of requests maxing at 4 per second is not actually a problem.
People complaining about fedi hammering their servers are probably using PHP or Python.
not the parent commentor, but I've previously heard complaints about how each server separately does a fetch of images and html for previews.
happened to me too!
Ok, on the one hand to have one bot hitting your website to populate some kind of slop newsletter is pretty annoying.
On the other hand, “multiple times per minute” doesn’t seem that high? Even if it was 60x per min that’s fully within the limits of what a 2015 laptop in the closet can handle.
yeah, it was a wake-up call that I needed to make certain parts of my application faster, and I've been working on it for the past few weeks
Once every 24 hours seems sufficient for the stated purpose.
Heh. I posted a GitHub project Show HN, a couple of days ago, and it never even made it out of new, but it got me some bots.
That's a big reason that I never post any of my apps here.
Welcome to hell, kid...
I’ve been enjoying https://www.unslop.news/ after it was shared here.
HN has become unbearably AI slop oriented over the last couple of years. It’s nice to see something else.
Glad you're liking unslop.news :) If you ever find any bugs or want anything added, it's open source so feel free to open a ticket https://github.com/leiDnedyA/hn-without-ai
I _also_ posted to HN and made it to #1[0]; the effect this had on my CPU/network traffic on a 5/EUR month VPS over at Hetzner was in the single digit percentages.
As per the wiki page for the C10k problem[1] where it says "_was_ [emphasis mine] the problem of optimizing computer networking stacks to handle a large number of clients at the same time". This was solved over a decade ago...
Now get off my lawn!
(TBF I think the author acknowledges this)
[0] https://news.ycombinator.com/item?id=43725815
[1] https://en.wikipedia.org/wiki/C10k_problem
The c10k problem is a different problem. It relates to simultaneous open TCP connections at a server. It would take a very very large number of crawler bots to force open 10k connections, I suspect.
I know it's a different problem but presumably there's at least some relationship/order of magnitude being expressed here...oh well
> It would take a very very large number of crawler bots to force open 10k connections, I suspect.
You can do that with any system made in the past 15, maybe 20 years.
From the wikipedia page:
"As of 2025, the problem has long since been solved, with the number of possible connections to a single computer being in the millions."