It's just a poor point. Cats are a bad analogy anyhow because there is no cat language. You can't actually speak to them or have a drawn out discussion. But you can with humans and many have inspired thousands or even millions to action doing just that, no super-intelligence required.
I really have wanted to do this but Fossil lacks MFA, OIDC, and of course CI/CD. Maybe there's a way to get all three in it but Idk. I know for OIDC you could in theory just use a reverse proxy to do it but then you have to get Fossil to respect it and not just ask you to login again.
Agreed, that's a huge turn off for me, and I thought this would genuinely be fascinating. I'm not a physics expert but I love reading about interesting things like this, but I can't stand this surface-level "well I in theory could be an expert on this topic but nobody knows because the machine removed all of the nuance and now it's shallow AI writing" style of writing.
I primarily use Incus for all container stuff, not Docker. Is problematic if I want to e.g. use a docker-compose file, but I (think) it protects against these things because incus allows me to create a vm and not a container if I really need that level of isolation.
If only there was a language which allowed one to express instructions for a computer to execute which was nearly unambiguous, precise, deterministic, and containerized such that the computer would do exactly what you told it to.
...
Oh wait.
Yes, the above was referring to programming languages. Which is what prompts are, essentially. It's just a different (and more verbose) way of instructing the computer on what to do. It also has a solution space of infinity and is ambiguous enough that there is no way to secure it because there are infinite combinations of saying anything imaginable. All prompt injections do is prove this point, over and over and over again, and "prompting" an LLM is just reverse-engineering programming languages in the worst possible way. I suspect that we will eventually have no other choice but to revert to using programming languages because they are the only way to get the kind of protections that people are trying to come up with with all these containerization and virtualization systems (which inevitably fail).
You make a fair and valid point about prompts, but you're ignoring the fact that writing code that's truly secure is also virtually impossible. The stack of layers that an attacker can target range from your own code, to library code (Heartbleed), container escape (maskedPaths abuse), OS (Dark Sword, Ghost Tap), hardware (Spectre, Rowhammer), etc. Security is really hard. Fortunately exploiting these things is also hard.
The belief that something is more likely to be secure because it's code instead of a prompt is likely only avoiding one particular type of attack. That's a win, but you probably shouldn't think of it as meaning your code is actually secure.
The problem is what is the alternative? I'm (not) defending them or this practice by any measure, but we all know what happens if you just open your site up without these, especially with AI bots which hammer servers and are in effect a legalized DDoS system. I've hated CAPTCHAs ever since I first encountered them and I can't wait for them to just finally die a permanent death, but I also don't know how we solve the "how do you identify a human and a bot" in a way which doesn't require server admins to have extremely beefy servers or similar setups to handle the extra load. I'm not going to do the "there HAS to be a way thing" either because, for all I know, this could just be one of those impossible-to-solve problems.
> we all know what happens if you just open your site up without these, especially with AI bots which hammer servers and are in effect a legalized DDoS system
No, we don't know. I honestly do not understand the problem. I run websites, both static and non-static. Granted, my sites aren't exactly the most popular internet go-to destinations, but I should be seeing this DDoS too, right?
I do see lots of requests. Nothing that any modern system can't handle. Computers are stupid fast these days. Unless you are doing something unreasonable, it's really hard to even notice this "extra load".
I understand there are sites for whom this causes problems, but I think these are rare and could be optimized not to do unreasonable things.
I think too many people are annoyed by AI companies (arguably understandable position), look at their logs and speak of "hammering", "DDoS" and "extra load", while in reality it doesn't matter much.
We do know, just ask anyone who runs a more popular site or does anything where abuse can be monetized (shopping, reviews, etc.). Avoiding that due to obscurity isn’t an answer because it’s saying you’re safe until something, possibly outside of your control, causes the bots to descend and give you an extra 500M requests with no chance of revenue.
I’m with OP: I don’t like this but the alternatives all look like the death of the open web.
The person you're responding to already said they ran a modestly sized site. What actual scale opens one up to abuse? If only the top 1% of sites need it, then it seems silly to say "everyone" needs it.
Stack Overflow was outside of the Cloudflare network for years, and anti-abuse was maybe 3 or 4 full-time jobs – much of which still needs to be done, because Cloudflare's anti-bot protection hasn't actually stopped it. Most UGC sites are not as big as Stack Overflow was at its peak.
I'm referring specifically to the activities of Charcoal (https://charcoal-se.org/) and their Stack Exchange staff counterparts, taken together. This is about large-scale platform abuse, of the sort that Cloudflare is alleged to prevent (but doesn't, really), not the more mundane (and laborious) task of manual quality control.
errr... so anything related to UGC now has a lower bound of 3-4 FTE? Sure, I'll hire a team of content moderators next time I think about putting a comment form under my blog...
Please read their last sentence again and think about how much it understates the difference between stack overflow in its prime and a normal website. Also the "much of which still needs to be done".
Yes? Cloudflare doesn't replace moderators. At all. It only allegedly filters bot generated content, it doesn't filter user generated content and doesn't even intend to.
It might depend on the tech stack. I run a small niche website but it has PHP and a database (MediaWiki/PHPBB) and without Cloudflare I'd estimate I'd need to spend several hundred dollars a month to handle the traffic. Traffic used to be tens of thousands of requests a day. AI has increased that to between 400k and 3M requests per day but it's not a smooth distribution. This is with bot fight mode on that greatly reduces traffic.
I adopted Cloudflare because it was getting DDoSed by the AI crawlers. I'm pretty sure all of them are vibe coding their crawlers and don't bother adding rate limiting as a requirement.
That was my point. I was trying to be gentle by mentioning "unreasonable" things, but seriously — how did we get to the point where less than 6 requests per second (that's 500k requests per day) is considered a DDoS?
I've spent some effort on optimizing my sites, but most of the effort was focused on avoiding unreasonable (stupid) work. Do I need a session for every request? No, I don't! Do I need a database fetch for every access to my homepage? No, I don't! Is it a problem to actually load all of my static content in all supported languages (24) into memory and serve it from memory? No, it isn't!
I use Clojure behind nginx on the server for my sites. Oh, and I also pre-compress all static assets to Brotli, so anything that handles brotli gets a static file served directly from nginx. I also use immutable assets with unlimited caching semantics.
Really — the problem is that we've grown lax and our software has become bloated, slow, and with unreasonable code paths. If every page fetch does 12 database accesses and runs through a slow interpreter, that is surely going to be a problem.
That's the traffic after rate limiting controls and bot fight mode. It's 3-4 million requests per day without bot fight mode and just rate limits. And as I said it's not a smooth distribution. Plus the requests are almost never for pages in cache. It's always stuff like loading all the message threads from the year 2000 or loading up the details of every page edit ever made to a wiki page.
If it was more static content it'd be easier, it's really the db being a bottle neck in a dynamic site.
Yes, the software could be better optimized but then I'd have to own the development of it. There is no reason a niche website should be getting millions of requests per day.
I second this. My website exposes a cgit and 99% of the traffic now is AI scraping the sources, but the load is nowhere near DoS territory. And this is running on the cheapest VPS I could find.
Not saying I'm not annoyed by the scraping; I am looking to block them, but I'm also not going to put the site behind the gatekeeper. If anything, Cloudflare must love AI scraping now for the same reason AV companies love malware.
Now, if you are running a PHP stack...yeah, maybe that's the problem right there.
Is there actually any plausible theory why "AI" would repeatedly scrape the same sites? Are there that many competing, completely independent AI labs? Is it cheaper to repeatedly scrape than to buffer the scraped data locally? (I find it very hard to imagine that it's easier to deal with changing/disappearing content than it is to stand up such a cache.)
> 99% of the traffic now is AI scraping the sources
I wonder if we should stop fighting this and instead create an API specifically for this purpose? Or, a central repository that you could send your data to and say to anyone wanting to scrape, "safe yourself some time and just get my data from this other place"
The thing though is that they are extremely idiotic. They are constantly, recurringly, scanning the same files, I suppose out of FOMO that a line might have changed. I don't know what a special API solves, especially because HTTP already has etags to save you from re-downloading the whole damn file over again. But these bots don't care. The extent to which they don't care is such that, after I temporarily took cgit down for kicks, they'd get 404s and still repeatedly ask for the sames files days on end.
The PHP stack isn't even the problem, it's having unauthenticated requests getting past the cache in the first place, something that most sites should be able to prevent.
If you're in any way semi-popular and a decent size, you're gonna get hammered. PortableApps.com was partially offline for weeks due to China-based AI scrapers. You block the useragent, they start hitting you with another one from the same IP in the same way. You block the IP, they switch to another. You block the subnet, they use another. At one point it was nearly a thousand different IPs from around China hammering away. For all intents and purposes, a DDoS. This wasn't a little "extra load", this was load that was thousands of times beyond what our legitimate userbase was using.
And if you're thinking about blocking all of China, while this particular AI bot didn't use them, a bunch of other ones I've encountered use VPNs and hacked clients worldwide.
Consider yourself lucky. But don't let yourself fall into the trap of thinking it's a nonissue for everyone else until it happens to you.
People shouldn't have to be experts or provision a larger server to run a UGC service that can withstand the sort of 30x more traffic I'm seeing from AI bots. Or rather, you didn't render the argument for why they should have to do that if they can just use CloudFlare's free tier.
Either way, it's easy to have all the answers when you've never had the problem.
Has anyone pointed an AI scraper at your server at all? Unless your website appears in search engine listings I don't think the AI scrapers will slam it. My server has never been hit by them but my server is also practically unknown. All of this said, I'm not going to claim that server loads can handle it because many sysadmins have claimed otherwise, and I would like to think that their claims are reliable.
As soon as you get your TLS certificate you get bombarded with scraping. You don't need someone to "point a scraper at you".
What matters most is usually how much there is to scrape. If you have like 5 pages that's nothing. For forum like websites where each thread, each user profile, etc. gets scraped that's when traffic increases. I just let them have at it with no issues though, computers are fast.
That's really weird. My experience is quite different: I have several subdomains and all of them have TLS certs and I haven't (yet) seen this (thankfully). Either that, or my server is masking it. The weird thing is that my server is an OVH dedicated box that doesn't exactly have top-tier specs, so I have no idea what's going on there. Very weird indeed.
I mean... It may be that most of the things I run aren't really scrape-able. I run Matrix (which requires authentication), an XWiki instance, Zulip, Terraria, Forgejo, Nextcloud, a Mastodon server... Most of those require auth behind my Kanidm instance to actually do anything. Well and most of them have APIs that are much better than "scrape the universe".
They showed up when the AI money did. The evidence is circumstantial, but… some of them are remarkably well engineered (from a “how difficult is it to identify this traffic” perspective, in a way that never existed before (I have been running a quite sizeable site for 8 years, over 200k registered users, and you don’t need to register to use 99% of it).
I run a quite large website and there are a few patterns.
The usage is extremely quick, and follows easy-to-spot patterns. We noticed a spike in bounce rate.
They never come from Google, and the bad programmed ones just crawl several pages at a time, faster than a user could do.
Then there's the crazy spikes in visits from specific countries, pretty much scraping the entire content. Often from pools of IPs. In some cases had 30% unexplained (meaning: it wasn't viral or a marketing campaign) random sustained increases in traffic.
There's also the fact they don't interact with the complicated widgets, so zero XHR requests other than analytics pings.
They also don't cause spikes in Google Analytics, so I assume it's blocked, but they show up in logs and in the internal analytics.
It's not enough to DDOS the website at all, but it's a lot of noise in statistics that we gotta learn to filter.
> They never come from Google, and the bad programmed ones just crawl several pages at a time, faster than a user could do.
I’ve triggered this kind of “bot protection” right here on Hacker News many times. I did that by having a bunch of Hacker News pages open and then closing and reopening my browser. I’ve also triggered it by opening a bunch of links in the background too quickly. I’ve also triggered it by reading the article, then clicking back and upvoting/favouriting too quickly. I’m also located in Singapore, which people have started to advocate for blocking here recently.
A single non-bot legitimate user can easily trigger these kinds of heuristics just by using the site in a way you don’t expect. This can affect some users disproportionately more than others, e.g. disabled people who need to use assistive technology.
It's circumstantial evidence, but Occam's Razor also applies.
It's not a hostile DOS in the traditional sense (I've mitigated a few of those) - no "pay us to make it stop", no pattern to the requests other than "fetch every unique URL a few times".
It wasn't happening until financial incentives to gather large datasets for AI training appeared.
Bad actors (using residential proxies & claiming to be a real browser) mostly showed up after folk started blocking ones that identified themselves as AI scrapers.
It's obvious to blame AI training because there's a shortage of better explanations. Who else would be paying for these (expensive) residential botnets, only to use them to (eg) web-scrape wikipedia (which offers free downloads of its content in a structured format)?
The simplest explanation of the technical behavior is "a bot coded to follow every link it sees & save the results", and the simplest explanation of the motive to run such a bot is "to train a large language model".
Exactly. They (and most of all, Big G) stand to profit greatly from this browser discrimination. What better than to make more sites use them by launching DDoS attacks in the name of "AI scraping".
A small, single EU country focused non-static e-commerce, with proper robots.txt instructions that worked perfectly well in the search & co bots -only "era" with rate limiting for nginx/php-fpm setup - is kinda struggling without CF to handle 15000 requests per 15 minutes, coming from Chrome "users" from IPv6. Best so far was an avg. server load in htop = 40 on an 8-core server x_x
That's 16.6rps. A single guy holding the F5 key on chrome can generate that much traffic and take down your website. That kind of performance was never acceptable.
People will always reframe their request numbers to avoid stating their pitiful requests per second numbers, it's hilarious. "This thing is handling hundreds of thousands of requests per day!" Like cool, you're barely making it double digit requests per second.
Maybe a plain WordPress install. Run something like WooCommerce and install a bunch of plugins to get the functionality that WordPress and WooCommerce should have built-in, and suddenly a cheap VPS can only handle 2 or 3 requests per second.
It's phenomenal how inefficient the WordPress/WooCommerce stack is.
Though the main issue I'm seeing is credit card testing, not scraping.
And I'm ideologically opposed to using a CDN (because it shouldn't be needed for such a small site!) so it's somewhat a self-inflicted problem...
"Security" plugins are also HUGE problem here, most of them turns "few cached DB SELECTs" (or static file read if you use caching plugin) into now a bunch of inserts, just to log/analyze "offender" IP and maybe block it, in many cases turning "blocking offender" to be more costly that would be serving the page without the security plugin
You can calculate traffic stats for a day by IPs/subnets and probably bots will stand out. If they are using IPv6 you can figure out the ASN and block it completely.
You get downvoted for these opinions but I agree. Most people that complain that their servers get hammered by AI bots are those that run very unoptimized servers that can only handle like 100 rps. I've never had any issues with any of my moderately optimized websites. A $10 VPS can handle sooo much traffic.
I think people get annoyed when it's suggested they spend time optimising or even re-writing their websites to handle high traffic loads just to cater to AI bots ripping their content.
It's also not always easy to do. I run a small wiki which is fairly optimised, nearly every page manages at least ~3k rps on a small VPS. The only exception is the diff page which is ~150 rps. Optimising that while still giving good output isn't that easy, but the wiki doesn't have many users so that would be fine if it wasn't for the AI bots.
The AI bots ignore robots.txt and were initially hitting the site with ~1k rps crawling every combination. Even that would be manageable as there's currently ~150,000 combinations, except they kept re-crawling the whole lot each day. The server could manage it but it was a massive waste of resources.
They were using residential IPs and only sending 1 request from each IP making it impossible to block. In the end I gave up and put a Cloudflare challenge in front of it. I don't want to use Cloudflare but the alternative is forcing users to login to view diffs or remove them entirely.
What I do is have more strict rate limits for non logged in users. You tell them to log in if they hit the rate limit. For non logged in users, you have a rate limit not just for IP, but also for /24 and /16. Forget about IPv6, IPv4 scarcity is a feature not a bug.
The bot I had was using unique IPs for each request. Some were from cloud providers but most were just random residential ISPs. I couldn't see any obvious connections so rate limiting would've had to be a global rate limit.
Each IP only makes ~1 request though so easy to detect after the fact.
I guess they will run out of IPs at some point so maybe if I had logged each one forever and shown a challenge only to them, it would have fixed it eventually. Just depends how big their pool of IPs is.
You were getting 1k rps, and each request was from an unique IP? So after an hour you got hit by 3.6M different IPs? And all from uncorrelated /16s? That seems hard to believe. Not that I don't believe you, it's just hard for me to grasp that whoever was scraping you had such a large and distributed swarm.
This is called rotating residential proxy service. You can buy it off grey market sites that are probably getting it from botnet operators. It costs about $2-$5 per GB.
Curious, but how do the bots figure out the combinations? Or do you have links to the diffs from other sites? I assume the diff takes two files in query parameters or something.
There really isn't a good reason for a wiki (or git host) to provide diffs between arbitrary revisions to unauthenticated users. Limit it to diffs compared to previous (which can be cached) and this problem goes away.
In any case, such labyrinths of expensive dynamically generated pages are no excuse for subjecting people requesting the start page to bot checks.
I managed to solve my scraper problems without optimizing much, but if I had to optimize I think the only option might be "don't use mediawiki" and that's an extremely obnoxious solution. Though maybe I could get there by throttling specific kinds of pages.
Same. Tritium and the blog have done stents on the front page here and high traffic subreddits and that plus bots has never been a problem. UX could be improved through a CDN but even that isn’t worth the trade-off for us at the moment.
I don't think it's just privacy, it also increasingly turns the web itself into a walled garden. The end result is that websites can only ever be accessed by "approved" clients - the latest Chrome, Edge, Safari and if you're lucky Firefox - and nothing else.
I haven't ever noticed Cloudflare having any issues on Firefox, so presumably that implies any unilateral actions in web standards have been worked around by CF to provide the service to Firefox as well.
I'm pretty frequently blocked by Cloudflare when I use Firefox on OpenBSD -- apparently it's too suspicious of a combination for their liking, or something. Even on Linux I've occasionally had issues. I've had to email site operators to ask them to change their configuration so I can actually be a customer of their business.
I think there's some chance we get a "proof of purchase" system where there is some entity that takes a $10 payment to give out a unique identity token that you need to present to visit most sites. if you have a revocation process for ones used for bad actors, it seems like it would work pretty well.
If the bad guys also had to pay $50/month/IP it would probably work.
The bad guys don't pay that much. And sometimes the bad guys actually use the IPs of other people (botnets on residential IPs) and don't pay anything at all.
They pay something. You can get a few ten cents per gigabyte for a voluntary proxy right now. I've never tried it long enough to get a minimum payout, so could be a scam for all I know (or maybe the minimum payout is the scam).
What would stop you offering someone a few tens of cents per GB to borrow any other token barrier you put up?
> we all know what happens if you just open your site up without these, especially with AI bots which hammer servers and are in effect a legalized DDoS system
So delegalize it. Strip searching everyone to paper over the fact that the societal contract has been broken only delays that.
You can easily calculate which IPs/networks bots are using by looking at where most traffic comes from and who requests lot of pages with non-human speed.
Each IP address is either from a residential proxy network, or from AWS / GCP / DigitalOcean. And each IP requests at human speed. 1000 of them are an issue though.
The most plausible near-term path is probably micropayments embedded invisibly in AI agents. Your agent that has learned what you value and can make a reasonable decision to allow a micropayment for certain content pays on your behalf without requiring a conscious decision each time, eliminating the mental transaction cost problem entirely. It's the mental transaction cost that arguably led to the failure of the micro payment model back in the early 2000s.
Although the cynical part of me says that this will result in malicious actors trying to trick agents into giving out a bunch of micro payments. There are counter defenses that can help detect and compensate for that, but perhaps the best we will be able to do is prompt user with the default agent recommendation.
It may always happen but it would happen less if we updated patent laws to fine people who filed invalid patents or enforced some kind of similar punishment. If you file a patent, it's up to you to verify that your patent is actually valid, and the courts shouldn't have to do that legwork for you. It also doesn't help that the patent office/components of governments don't review patents as thoroughly as they used to. Same with trademarks.
I generally don't like the current patent law but it sounds a bit off to pay the government & wait for them to review your patent claim and then get fined by the government when both of you were wrong about it. There are already processes to additionally fine a company bringing about a truly frivolous patent lawsuit, it's just rare because usually it's not so cut and dry as we'd like it to be.
I mean my idea isn't the only one in that solution space. My reasoning was to ensure that the government actually reviewed the patent and ensured it was valid instead of rubber stamping it. Or, even better, the filer of the patent application would do that. Although the best is probably to make software unpatentable anyway.
> In the 1980s, when IBM accused Sun of violating seven patents, Sun examined the patents and argued that IBM didn't have a case. The reply of IBM's lawyers was "maybe you don't infringe these seven patents. But we have 10,000 U.S. patents. Do you really want us to go back to Armonk [IBM headquarters in New York] and find seven patents you do infringe? Or do you want to make this easy and just pay us $20 million?" And Sun paid out.[4]
Ban on software patent is more than warranted, only trolls themselves don't like such idea obviously.
But what also needed is persecution these patent trolls for racket like any mob racketeers. There were attempts of persecuting them this way in the past, but they should be renewed.
reply