A tiny mouse, a hacker.

See here for an introduction, and my link tree for socials.

  • 0 Posts
  • 42 Comments
Joined 3 years ago
cake
Cake day: December 24th, 2023

help-circle
  • Is that human or AI.

    As far as I know, it’s mostly incredibly underpaid human labor.

    Is this possible?

    Possible? Most certainly. The trouble is, we have absolutely no way to know, so the best we can do is… guess, hope, and pray. Preparing even one page of plausible looking text is hard - there’s a reason LLMs require powerful GPUs. Generating an infinite amount of plausible looking pages is even harder. That’s why I believe it is not worth the effort to poison them - it costs too much on our side, and we can’t verify if it works.

    On top of that, the more plausible it looks, the more likely it is to lead a human visitor astray, should they accidentally find themselves in the maze. Right now, if a human looks at the iocaine output, it is clearly garbage. The bots don’t care, and the programs and/or people who need to verify it, will either have to go through a massive trove of junk, or throw away everything scraped from the entire domain (in both cases, mission accomplished: LLMs didn’t get to the juicy data they wanted).


  • Ah I was wondering if you’d find this thread

    I did! It was lucky I found it so soon, I usually do a search once or twice a week - and it just happened I did one of those yesterday.

    previously I kept running into articles about princess bride

    That’s normal, iocaine’s name is from the Princess Bride, after all - and the book and the film are much better known than my software (as they should be).

    Unfortunately I have nothing to host on the internet but I wonder it could be paired with something that traps the clanker in the maze with a proof of work in between each page. That way they are ingesting poison and working for it. I get a lot of joy from the idea of a swarm of bots stuck in an infinite maze.

    The ingested poison doesn’t do much, it’ll get discarded before training. So if you have nothing to put behind iocaine, I don’t think it’s worth it to run it. It can generate endless amounts of garbage, yeah, and will tie up some crawler resources too, yup, but it also costs you bandwidth and CPU & RAM you otherwise wouldn’t need to spend (since there’s nothing behind it, there’s no “more expensive” endpoint iocaine is taking a hit for). I don’t think that’s a good trade. As for Proof of Work: most bots won’t bother, and for the few that do… the cost/benefit ratio is still bad if there’s nothing behind iocaine.



  • I noticed you started another rewrite of Iocaine for version 4.x. May I ask about plans for version four compared to version three?

    I wrote a bit about the plans over on my blog. The major changes in a nutshell are:

    • Nam-Shub of Enki elevated to “default script” status
      • It also becomes a module library, and will be available in both Roto and Lua
    • Roto upgraded from 0.9 to 0.13 (bringing a whole lot of nice language features, more performance, and less code I need to write)
    • Half the dependencies axed.
    • Plugins for Caddy, nginx, and Apache, embedding iocaine.
    • Much, much, much better error reporting (see an example of an error here)
    • Throttling & rate limiting

    How well does Iocaine integrate with OpenRC? I don’t use SystemD much, as I tend to run Artix.

    Pretty well, I think? There’s an OpenRC init script in the repo. I haven’t tested it myself, I currently use systemd on Linux, and NetBSD’s init system or dinit on NetBSD. The only systemd-specific thing in iocaine is that it notifies systemd when it finished starting up - and that’s going to go away in iocaine 4, because I’m cutting dependencies, and this tiny thing isn’t worth a dependency.


  • FWIW, while iocaine does generate copious amounts of garbage, it is never going to poison LLM training data: it’s trivial to filter iocaine-generated garbage out pre-training. The garbage is there for two reasons:

    1. To serve the crawlers something.
    2. To poison the crawling queue.

    Not the models, the crawling queue. Place an identifiable substring into every generated URL, and you can catch any crawler that works off of a queue collected by cheaper crawlers.

    It’s not really a poisoning tool - it originally started with that goal, but that goal has changed a loooong time ago -, it’s more bot detection & mitigation nowadays, the garbage is just one of the tools it achieves this goal with.

    But while I’m here, I’d like to mention Alun Jones’ fakejpeg (or my rust port): a library that can help you generate structurally legit-ish JPEGs, files that most decoders will happily display, but with entirely random content. Generating them is very cheap. How effective it is in poisoning? I don’t know. But I find great delight in looking at the random noise.

    Hello Lemmy





  • Way back when, my Dad always told me to have at least three copies of everything, because one backup is no backup. So I have a 1TB Storage Box at Hetzner (€4/month, in a different datacenter than my VPSes) I push backups to with Restic. I also backup to a local machine (also with restic), and to my Dad’s NAS (you guessed, with restic), and to my old Debian server until I decommission it (also with restic).

    All legs have different retention settings, based on how much space I have at a particular backup site.





  • Uh, I beg to differ. I’ve been blocking most of them for the past year, using essentially three ifs in a trenchcoat.

    Bullshit user agents are taken care of by checking headers other than user-agent: if they say they’re Chrome/ or Firefox/, check if they sent sec-fetch-mode. Didn’t? That’s very likely a crawler (and the handful of false positives are easy to make an exception for). For residential proxies, the same applies. For crawlers that piggy-back on Chrome, they usually crawl an URL queue, so if you poison their queue, you can catch those too.

    At this point, out of ~100 million requests / day, I’m firewalling ~98 million off. Out of the remaining 2 million, ~90% of them gets served garbage to continue poisoning the URL queues. I can serve the rest on a potato, even if some of them are crawlers.

    I did have a few people contact me about false positives, but those were very, very few (and also very easy to address). Very little CPU, RAM or bandwidth required, the vast majority of bots caught, negligible false positives. Deploying the solution isn’t trivial (yet), but it also isn’t hard either.



  • I’m running Tang on a VPS, outside of my homelab. Servers in my homelab set up networking and a dedicated WireGuard tunnel to the VPS from initrd, to be able to talk to Tang, to help unlock the filesystem. The WireGuard tunnel is only allowed from my home ISP’s ASN. So if anyone picks up all my equipment from my homelab and walks away with them, they will not be able to boot them up, unless they connect from my ISP’s ASN (good luck), or know the passphrase.

    Additionally, some of my homelab computers that support TPM also have a TPM pin, so walking away with the disk only, and connecting from my ISP’s ASN would still not be enough. This is rather pointless, anyone who walks away with the disk only will likely take the entire computer instead. But it was fun setting it up.

    In the not so distant future, I’ll update this setup to use Shamir Secret Sharing more, where I’ll have three pins: my VPS (via Wireguard), a small computer somewhere else in my apartment, and a third at a neighbour (+ TPM on supporting computers).