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
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.
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.
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).
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.
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
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 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'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?
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.
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.
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.
> 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
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.
>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.
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.
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
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.
> 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 aswell silently drop the packets if they don't adapt.
> 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?
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?
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.
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).
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.
I ran into similar problems. My analytics became so useless that I threw them out.
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
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).
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.
What is that improved strategy? :pray:
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
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 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.
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 really recommend blocking all traffic from anything on https://github.com/borestad/blocklist-abuseipdb
That is primarily why websites only TLS-negotiates for server-side-only ChaCha-Poly.
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?
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.
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
> one specific user. I won't name names
Who is?!
haha im not naming names
Does the country end with an "a" with 5 - 6 characters long?
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?
> 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
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 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.
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.
Oh, as soon as you add tls to your subdomain, the bots come. It’s instant.
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.
Same with me, every-time I got 1st page I got a lot of robots
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.
not the parent commentor, but I've previously heard complaints about how each server separately does a fetch of images and html for previews.
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...
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.
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.