> because apparently what we have to offer is worth spending a ton of cycles to calculate the Anubis challenge.
This statement holds the core misapprehension behind Anubis. It’s not a ton of cycles. There is no difficulty setting that would be inconvenient for bots but usable for humans on mobile devices.
I noticed the other day that lists.ffmpeg.org had moved to Anubis difficulty level 6, which takes ~180sec for my iPhone 17 to solve at ~100KH/s, making the site unusable. So I spent ~10 minutes vibe coding a safari extension with a native bridge to an optimized C kernel using ARM SHA256H* instructions that can do 200+ MH/s on the same device. This solves Anubis difficulty level 6 in a handful of milliseconds.
Given the numbers and capabilities involved (a single $5K ASIC miner yields 200TH/s, a million times more hash rate than my optimized kernel running on an iPhone), I don’t see how proof of work could possibly be a sustainable strategy to keep bots out without ruining human user experience. It’s an arms race that can’t be won.
Edit: I encourage you to try this yourself. Here's a sample prompt that ought to one-shot the task:
> Build an iOS Safari Web Extension that accelerates Anubis proof-of-work using a native C ARM64 SHA-256 kernel. Precompute the invariant 128-byte challenge prefix, search fixed-width decimal nonces with ARM SHA-2 intrinsics and two worker threads, and target difficulty-6 solves under one second. Relay challenges from a Safari content script through the background service worker to native code, then submit the valid nonce/hash through Anubis’s normal pass-challenge endpoint. Include a deterministic benchmark app, correctness tests against CryptoKit, bounded execution, and fallback to Anubis’s stock solver.
Isn't the issue here the three order of magnitude difference between the (presumably optimized) JS implementation and the optimized C kernel on your browser? If the two stay within an order of magnitude of each other Anubis is a perfectly viable and scalable solution. Bots want to open many orders of magnitude more pages than a normal user, so the resource investment for each single page matters a lot more to them
For reference, the challenge on lists.ffmpeg.org takes 8 seconds on Firefox on my three year old laptop CPU that has worse benchmark scores than the iPhone 17 (tbf, the laptop also cost less than an iPhone 17). 8 seconds doesn't run against thermal limitations, so I really don't see why Safari on a modern iPhone should be so slow at this
> I really don't see why Safari on a modern iPhone should be so slow at this
me neither, but I don't think it changes the argument. There's always going to be someone on a low-end device. Your adversaries already have superhuman coding ability and infinite patience. Why would you expect the long-term advantage to be with the defenders?
I highly doubt the adversaries actually care about kernel.org in particular.
It's more likely that they're trying to scrape the entire internet, and cgit is just unfortunate enough to have autogenerated pages, where each page seems to contain text entirely distinct from its neighbors. This probably bypasses the scrapers' heuristics, which normally prevent them from getting trapped in a maze of twisty little autogenerated pages, all alike.
The way you normally deal with datasets at that scale is to move postprocessing (particularly expensive postprocessing) as late in the pipeline as possible.
It’s more economical at a compute level, but not at the developer level. The moment you start customizing your crawler to use protocol X for site Y your scale story collapses.
It's a good point, but in practice it depends on how easy those customizations are to implement / maintain, and how much money and effort you save. At some point the compute cost can disrupt even the nicest scale story.
I think the path forward is that websites offer one path for humans, and another for scrapers. But the huge catch is the path for scrapers must be _genuinely_ and _reliably_ the more economical and scalable path (either through something like PoW arms races, or through fear of litigation). Otherwise they will continue to ignore instructions and intrude on the human path.
The temptation to offer different / inferior / limited content to scrapers will be too strong, so such solutions are doomed to fail.
This would likely work as "Cloudflare SideChannel", a (hypothetical) Cloudflare product that would let scrapers download the pages that humans actually visit, as they are added to the CF cache. It wouldn't work for the non-Cloudflare part of the internet where humans connect directly to the servers that have their content.
The moment we establish a standard for offering a "optimized for scrapers" version of a site, people who do not want to be scraped will weaponise that to serve junk to scrapers... and scrapers will subsequently refuse to use it.
Largely because they're residential botnets in places like Brazil (a real example from one of my sites that was crawled to near-destruction). Someone could probably do something about this, but it's out of reach for individual site owners.
How good are you personally at iteratively and sequentially jumping through the legal systems of dozens of countries over the course of years, interleaved with genuinely difficult technical investigation, to unmask successive onion-layers of identity in order to unmask one offender?
Oh, and it also only takes a few minutes to reconfigure everything and invalidate those years of legal and investigatory work.
The point of Anubis isn't to prevent bots from accessing a site, but rather to prevent dumb bots from accidentally DOSing a site.
If someone adds an Anubis solver to their spider, but then also makes it respect cache tags and stops crawling random old commits through a git-web interface, that's probably still mission accomplished.
But the tokens are IP-bound, and as the article mentions a crawler using a single IP is not so difficult to block.
A legitimate user solves it once, shows legitimate-use behavior, and retains their access for thousands of requests. A crawler solves a challenge, is determined to use bot-like behavior after a dozen requests, gets an IP ban, and needs to recycle their IP and start from scratch.
It is indeed worse for fly-by users, but with a website like Linux cgit those basically don't exist. It would be a completely different story for, say, someone's personal blog.
Now you have a session of sorts and can limit the requests for that client, right? They can be fast, just limited in volume - regular user isn't punished.
So, they can interleave different tokens and get stopless access at normal price?
Yes this needs to be tailored for Anubis, but that was the point: the moment it will become popular enough, scrapers will have no problems to bypass it.
Paying a second of CPU time is an inconvenience to a user but trivial for LLM scraping.
The GPU cost to pretrain on that page once will dwarf by 2-3 OOMs the CPU cost to compute Anubis, scrape and post process it. And you’re not going to just train on it once!
At best you’re creating a speed bump for wannabe players scraping with no real plan. The folks training models people use just do not care.
> If the two stay within an order of magnitude of each other Anubis is a perfectly viable and scalable solution.
Exactly, and that's even before accounting for more sophisticated PoW algorithms such as the one used by XMR that almost entirely eliminate any benefit from offloading to an accelerator.
No it's not, beacuse the people running the bots have access to the c implementation and GPUs, the arm implementation on the iphone might be fast, but is almost certainly slower than an implementation running on a 5090, let alone a b300.
It's like trying to keep billionaires out of a restaurant by charging high prices, but still keeping it affordable for the middle class, it doesn't work on a fundamental level.
The people trying to do the scrapping will always have more compute than your typical user.
> I don’t see how proof of work could possibly be a sustainable strategy to keep bots out without ruining human user experience.
Lots of work has been done to do just that by monero’s proof of work algorithm. They achieved ASIC and GPU resistance by making you generate entire programs (with quite a bit of RAM usage) and finding one that produces a desired output. Then you submit the seed used for the program (a series of programs actually but let’s not get into the details)
One possible way to do this: rather than simply burning hashes for nothing, Anubis can burn hashes to mine crypto. This way, the bots pay for the infrastructure they consume, at least in part.
Combine with heuristics like in the article, so you can crank up the difficulty only for likely bot traffic, and you've got a bot supported site.
CPU crypto mining stopped being viable over a decade ago. If you want a client to pay for the infra cost of their request, they're probably going to need to work for hours to do so.
> Anubis wants "slow for bots but fast for humans"
That's not true. The goal is to raise the cost for bulk scrapers to the extent possible while minimizing any inconvenience to regular users. It's not about being slow for bots but rather reducing the asymmetry of the problem (from the perspective of the server) by imposing a (minor) cost on anyone requesting a page.
That's a better way to put the goal. But does it actually raise the cost or reduce asymmetry in a meaningful way? If it makes crawling 0-5% slower then it's really not worth the annoyance.
I know having the Anubis javascript at all gets in the way of some bots. But is the calculation helping enough?
> does it actually raise the cost or reduce asymmetry in a meaningful way?
It's complicated. I think anubis currently only collects low hanging fruit so to speak but in practice that's the bulk of what was causing problems so it's good enough.
PoW in general absolutely yes - I think monero clearly demonstrates this. Last I checked it remained viable to mine monero on at least somewhat low end and at least somewhat outdated consumer CPUs. It seems to me that fact implies that a well constructed PoW scheme can in practice squash the impact of any performance differences to within at most a couple orders of magnitude.
Sure. Even if PoW in general is sound, when applied to scraping it's always going to be a matter of what the client is willing to "pay". It's entirely within the realm of possibility that a scraper could be willing to "outbid" the user of a 5 year old tablet.
But it's also true that charging anything is going to be expected to reduce server load, that the cheaper something is the more carelessly people will tend to make use of it, and that at least in general bot operators are expected to be more cost sensitive than regular visitors on account of scale.
To that last point consider that a regular visitor will generally have an idle CPU and multiple gigabytes of free RAM. Meanwhile a competent scraper will generally be optimizing request parallelism per cheap VPS while an incompetent one will just point and shoot (and proceed to blindly visit every cgit url). The latter is by far the most troublesome and conveniently also the most likely to be impacted by even a poorly implemented PoW scheme.
I keep seeing comments along the lines of anubis only working because it bounces bots running lightweight setups without js but stop and ask yourself why are they running such lightweight setups to begin with?
> why are they running such lightweight setups to begin with?
It's easier and still gives access to almost everything. I don't think CPU limitations are the main issue there. As far as computer resources, I'd bet RAM is the bottleneck to running the script, not cycles. That suggests a RAM-heavy hash might do better than SHA256, but still agrees with the idea that the PoW in Anubis is not the important part.
> To that last point consider that a regular visitor will generally have an idle CPU and multiple gigabytes of free RAM. Meanwhile a competent scraper will generally be optimizing request parallelism per cheap VPS while an incompetent one will just point and shoot (and proceed to blindly visit every cgit url). The latter is by far the most troublesome and conveniently also the most likely to be impacted by even a poorly implemented PoW scheme.
I considered that. But while a visitor is idle they need the result quickly and they face the full brunt of the challenge all at once. A scraper can be working on a thousand sites at once so the per-request impact of the challenges is a lot lower, and patiently working through challenges in the background usually reduces the real cost, potentially to almost zero.
And you'd better not need multiple gigabytes of RAM to visit your site, that'll freeze my computer.
The WebAssembly challenge (as of yet unreleased, it's been in dev for a year and I don't want to enable it by default until I know it won't blow up horribly) is bit-wise instead of nibble-wise, meaning that admins have better challenge granularity: https://anubis.techaro.lol/docs/admin/configuration/challeng...
What I'm hearing is that you spent a year designing a more complicated proof-of-work challenge rather than moving to a challenge system that actually impacts scrapers. Consider revisiting this decision!
> rather than moving to a challenge system that actually impacts scrapers
At this point isn't it basically auth-only?
Rant: (genuinely wondering too, and RFC, request for conversation) at this point don't we have Google, etc. basically doing Real World ID Verification, but without an open protocol backing, using it to corral users into their ecosystem and gather data, and leaving us without some open and distributed alternative like the (fairly well working it seems to me) Public Key infrastructure? (PKI)
I know Sam Altman (with Proof of Human or w/e) and others have suggested and also have their own alternative (which I think is basically what Google already does), but I think this should simply be built based on an open protocol and then hopefully not-too-large (and with a healthy competition) local companies could do it for you.
Maybe, but at least you can more easily assign responsibility in this case. If you soft-block or hard-block someone because they're in a botnet, you can get them to change (expire the old ones) their credentials and start with clean systems. Maybe they'll realize their TV is part of a botnet if they have to keep doing it, or their PC has malware.
Indeed, but then you can (for example) just basically blocklist this ID and share this blocklist with other servers (and likewise import blocklists) -- a blocklist curated by one or more NGOs or the sort would be ideal.
Fixing a malware infection should be like fixing up anything else broken around your house (clean things up, re-auth yourself with the ID verifier, done).
Also, bots cannot simply automate renewing the credentials of course, after expiry you should be required to physically (or in a trustworthy digital medium) re-authenticate yourself preventing any kind of mass attack, you'd need millions of people to perform authentication multiple times a day (by that point it seems safe to assume they'd rather fix the malware source in their personal devices).
I am actually working on this kind of shared reputation database: https://github.com/TecharoHQ/reputationdb (warning: very prototype, not documented, such bugs, wow). I just have more to do than I have free time at this point.
And if you're a bot operating at scale, you can use different devices for crawling and challenge solving.
Do you have an underpowered TV box that would need hours to solve difficulty 6? Do you also have a Windows box where somebody installed gta6_leak_full.iso? Great, have the TV send their challenges to the Windows box for solving.
If you're operating at scale, you probably have hyper-optimized Anubis kernels, E.G. compiled for the particular set of SIMD extensions that a given machine has. At that scale, getting 1% more juice out of your machine pool becomes worth the squeeze. In the worst case scenario, you can probably get some cloud boxes on AWS to do the challenge solving for you.
Out of curiosity I tried to load this on my iPhone 12 and it took about 8 and a half minutes and once it finally passed it said my browser didn't allow cookies and to enable them before trying again. I would have to go through the whole proof again. I get what they're doing but there definitely needs to be some work done on getting those times down on iphones.
Tough. The code isn’t worth enough to publish. It means nothing to me and I’d have to figure out how to do it pseudonymously. I only have the one GitHub account.
It's kinda interesting to see simultaneously the 'holy shit' response to the OpenAI / HuggingFace incident, and then the griping about Fable's controls regarding this.
It does work, at least in my experience. The challenge may be cheap for your n-of-1 but keep in mind the scraper bots are crawling number of domains with very high concurrency.
And without control of their source IP addresses. They use a "rotating proxy" meta-proxy which sends each request through a different proxy, which invalidates the Anubis response and triggers a new challenge.
Okay. That's good enough. The problem people using Anubis have is not "I'm being targeted for a DDoS" and is actually "my site is being incidentally DDoS'd because scrapers are DDoSing the entire internet". A solution to the latter is what people need. Every single one of these threads comes up with hypotheticals about how useless Anubis is and how it can't possibly work and yet it does work for the real-world problem that real-world webmasters are dealing with.
Right: It's a threat-modeling mistake to start lumping different groups with different goals and different operational constraints together.
Sure, the cheap lock I added to the outside shed's door-latch is terrible and can be shimmed open in two seconds with a bit of flat metal... but the important thing right now is that it keeps the raccoons out. If there's a new threat to the gardening tools or the bikes, then I have a different problem and I'll consider something else.
Most of the problem in this case is that the scrapers are being incredibly lazy and not investing even minimal effort in their approach. If this forces them to revisit how they go about things then mission accomplished.
One natural counter to this is to dynamically alter the computing algorithm per request (variants of Anubis?) so that the scrapers would need to rewrite their accelerator each time or just use the unoptimized JS version...
anubis was a joke on arrival. PoW for this purpose is plain stupid.
the only feasible CAPTCHA at this moment is probably hCaptcha. they began to offer animated challenges. it's probably the best you can do, using an LLM to solve them would be a bit expensive for now.
As the article mentions: Anubis blocked virtually all bots for a couple of months, and even now it is still blocking 67% of requests. Without Anubis they'd have to double the number of CPU cores, solely to serve crawler traffic!
It's an arms race. The best solution is the one you can plug in right now - even if it is stupid. hCaptcha might have a better blocking rate at the moment, but it's also $0.99/1K evals, and that "animated challenge" will be impossible to solve for visually-impaired users: for many people that cost is significantly higher than spinning up an extra server or two to feed the crawlers.
My iPad Pro M2 takes around 30 seconds to solve it in Safari, while my Mac Studio solves it in about 3 seconds using 449% logical CPUs. It really makes no sense. Interesting how much slower it is than native - could it be protections against things such as crypto mining?
I'm reminded of LUKS Argon2 implementation, which is much slower than userspace. Therefore a sufficiently difficult challenge that may take less than a second in userspace Linux can take 20 or more seconds at boot.
Memory-hard hash functions maybe? Like, you must dedicate 4gb of ram to compute the function. Not a problem for a one-off, but is a problem when reading lots of pages at once.
Or… the site will serve a random seed and the device must compute 4gb of pseudo-random data, then supply a value at a random server-demanded offset.
>It's sole purpose is to deny the bots you don't want on your site.
This is fundamentally impossible.
In the end only pay per crawl will work. Like x402 but maybe something similar.
In other words: not only client has to spend some kind of resource for access, but it has to be a transaction, server must get something from serving the request.
> Anubis is not a load optimizer. It's sole purpose is to deny the bots you don't want on your site.
No, this very article will tell you that the purpose of Anubis is to reduce load on the site. The purpose of denying bot requests is that you're strained by the load they impose.
Conversely, bots that pose no problems to you aren't "the bots you don't want on your site".
This is the basic misconception behind the Anubis design. In an anti-spam system, there's a user/abuser reward asymmetry. Users make single requests and get some constant reward V. Abusers make millions of requests and also get some constant reward V. Proof-of-work exploits the asymmetry: costs are, by a factor of ~millions, disproportionately imposed on abusers.
That's exactly not what happens with scrapers. Scrapers make millions of requests, yes. But each one of those requests produces an average (token-denominated) reward V'. There's no asymmetry at all: you're just adding a fixed cost to each reward, for both users and abusers.
That's before you even get to questions like "does the system actually work, or does a single solve get practically unbounded rewards".
> There's no asymmetry at all: you're just adding a fixed cost to each reward, for both users and abusers.
There is: users solve the PoW once, scrappers solve it every time they rotate IP addresses. Some of them change their IP address on every request, so they keep solving the PoW over and over.
That doesn't matter! Every time they successfully solve the challenge, they're getting the token reward. That's the missing asymmetry. In an anti-spam system, 99.999% of all transactions are unrewarded.
I have a small personal site with an exposed Git repository. I'm getting crawled from about 100000 residential IPs from several ASNs (mostly in Brazil). Each crawler makes a couple of requests per minute.
I suspect that's most of the value, and that the deterrent is about software construction and not about the proof-of-work, which I've been saying for the last year is a productive direction for Anubis to head towards. Instead, it's just sort of tormenting normal people, as a kind of performance of anger about AI.
Which: I get! If it's your site, perform all the anger you want.
OK, but that doesn't change anything. You have a large pool of IPs, each of which only needs to expend a handful of extra milliseconds of work to get unlimited access to the protected resource.
Even if you had to solve a L6 challenge for every request it's faster than the total RTT time of most servers. In other words not a meaningful barrier. And L6 is already a level which severely interferes with human usage of a site.
a $5000 ASIC-based device can hash SHA256 at 200TH/s or more.
And yet many report it works, at least for now, and the excess load due to scraping activity falls precipitously when Anubis or similar solutions are used. Maybe once this sort of challenge is used almost everywhere we'll see concerted attempts to get around it, but for now it is easier for them to just move on to another target.
> a $5000 ASIC-based device can hash SHA256 at 200TH/s or more
Peanuts for the big players, but many (almost all?) running smaller scale scraping operations are going to find $5000 rather prohibitive, and they are unlikely to be able to integrate it as they are probably running a “stock” scraper that they didn't write themselves.
You don’t need to spend $5000 to obtain the hash rate of a $5000 device on a rental basis. You may have heard of this thing called “the cloud”. Obtaining very high hash rates is effectively free, largely as a side effect of the crypto bust.
Not sure why anyone would characterize these scrapers im general as all being fly-by-night operations that don’t have two cents to scrape together.
A significant amount of the effect is everything outside the proof of work. Not the cost of hashing but the need to run the javascript that submits it.
If you've seen anyone post a comparison of crawl rate versus difficulty, I'd love to see it. There's probably some difference but I want to know how much of the overall effect it is.
I never said the approach was so stupid it could never work. I just shared how it’s annoying and how I worked around my annoyance.
An even stupider approach would work exactly as well or better, like a form saying “type the letter y in this box to continue”. The only benefit is as a road bump that makes the site in any way custom. The moment anyone you are defending against so much as looks at the mechanics of solving the challenge it completely falls apart.
The way they're internally implemented doesn't allow pinning an IP. They buy a rotating proxy service from a vendor, and don't get to choose their source IP.
It's not hard to test. Go to a page that demands PoW, change your IP and see what happens. I just did it. Spoiler: kernel.org asks for a new PoW.
If the source IP was an issue, you could do it other ways: for example, make the cookie rotate on every access, and insist there is a single stream of accesses.
Why are you and other defenders of the Anubis approach so fixated on this one specific limitation of a certain type of scraping architecture? It’s hardly an immutable characteristic.
You say “they” as if all scrapers are a monolithic group with the same constraints and goals. Part of the problem is the massive diversity.
Some of the asymmetry might be regained if anubis had thousands of variations of PoW algorithms, each different enough that they must be solved independently.
I wonder if AI might be able to come up with new PoW algorithms in a nightly CI job so every day is a different puzzle...
You can do this on yours. Just have the client and server add an extra "2" after the challenge key or something. A different client which extracts the challenge key and does its own processing will only generate invalid responses.
Cool, so then that invalidates the ASIC problem, right?
My earlier idea was to imagine that each day Anubis picks an entirely different problem-class. Ex: one day it is Sha256, the next it is prime factorization, the next it is twin-prime-finding, the next it is cracking elliptic curves, the next it is some kind of sorting / information theory problem...
All with the goal of adapting constantly so that scrapers have a harder time optimizing for the PoW problem (i.e. with Sha256 ASICs)
...That sounds like entropy? As in, the thing computers are bad at (truly random numbers) and /dev/urandom in your kernel already spits out an approximation of?
> I don’t see how proof of work could possibly be a sustainable strategy to keep bots out without ruining human user experience. It’s an arms race that can’t be won.
You yourself describe the solution.
Extensions or browser integrations of the Anubis task.
He didn’t describe a solution. He described a (crappy) workaround for humans. But the fact is that this cannot and will not stop bots. The people running bots can do the same, even faster.
I’m left wondering if we disagree about what the problem is.
The problem here is not merely that Anubis is inconvenient to humans. It’s that and also that it’s not very effective for blocking bots. Anything that makes it easier for humans to get past will also make it easier for bots to get past.
I mean, it works just fine for PoW-based cryptocurrencies (with "why are we doing this if there are much better solutions to the same problem" values of "fine") because there the game theory works out, which is "solve hashes to gain a probabilistic lottery ticket for a shared prize pool" having a higher EV than "solve hashes for the probabilistic chance of double spending some money, but if you don't pull it off you gain zero".
Anubis is just "solve hashes to get access to a valuable resources", and if the resource is at least valuable to whoever you're trying to deter as it is to legitimate users, you're not deterring anyone (and by extension, if it's more valuable to them, you're deterring your legitimate users instead).
I mention litecoin specifically, it was developed right after BTC ASICS started dominating, and the LTC goal was for hardness to reward consumer hardware rather than ASICS, it was briefly successful until LTC ASICS were developed.
That's actually weaponizable. You can deliberately insist on an inefficient algorithm to find bots that use a better algorithm to bypass the PoW. Legitimate users have zero reason to redo the challenge over and over again, meanwhile bots want to constantly redo the challenge because they don't want you to throttle the bot based on the reused session cookie. You can then crank up the difficulty up over and over again.
I noticed the other day that lists.ffmpeg.org had moved to Anubis difficulty level 6, which takes ~180sec for my iPhone 17 to solve at ~100KH/s, making the site unusable. So I spent ~10 minutes vibe coding a safari extension with a native bridge to an optimized C kernel using ARM SHA256H* instructions that can do 200+ MH/s on the same device. This solves Anubis difficulty level 6 in a handful of milliseconds.
Given the numbers and capabilities involved (a single $5K ASIC miner yields 200TH/s, a million times more hash rate than my optimized kernel running on an iPhone), I don’t see how proof of work could possibly be a sustainable strategy to keep bots out without ruining human user experience. It’s an arms race that can’t be won.
Edit: I encourage you to try this yourself. Here's a sample prompt that ought to one-shot the task: