[HN Gopher] Trapping misbehaving bots in an AI Labyrinth
       ___________________________________________________________________
        
       Trapping misbehaving bots in an AI Labyrinth
        
       Author : pabs3
       Score  : 76 points
       Date   : 2025-03-19 13:31 UTC (3 days ago)
        
 (HTM) web link (blog.cloudflare.com)
 (TXT) w3m dump (blog.cloudflare.com)
        
       | windhaven wrote:
       | From the post: "It is important to us that we don't generate
       | inaccurate content that contributes to the spread of
       | misinformation on the Internet, so the content we generate is
       | real and related to scientific facts, just not relevant or
       | proprietary to the site being crawled."
       | 
       | Okay, why should I care if a crawler that is clearly doing
       | something it shouldn't receives misinformation?
        
         | bithavoc wrote:
         | I guess if the crawlers can't actually see the trap,
         | misinformation would be attributed to your website in case
         | model responses expose content attribution tags to end users.
        
           | dns_snek wrote:
           | LLMs already make up citations, everyone would assume it was
           | just the model spewing nonsense citations.
        
         | mentalgear wrote:
         | That's actually a good strategy. It avoids adding more false
         | information in the infosphere while de-incentivising the
         | crawlers from returning to the site (since they don't find the
         | information they are looking for there).
        
       | d4rkn0d3z wrote:
       | "When we detect unauthorized crawling..."
       | 
       | How did you do that?
        
         | mog_dev wrote:
         | Simple, you add the trapped paths to robots.txt Well behaved
         | robots will not crawl them.
        
           | d4rkn0d3z wrote:
           | Nifty trick.
        
         | hombre_fatal wrote:
         | Just consider how you click around HN versus how your crawler
         | would behave if you wanted to crawl every page of HN starting
         | from the homepage.
        
       | theyknowitsxmas wrote:
       | The toggle doesn't work for me.
       | 
       | API Request Failed: PUT /api/v4/zones/xxx/bot_management (504)
        
       | visarga wrote:
       | Like burning down you house so squatters can't use it. A
       | poisoning attack that makes communication itself untrusted.
       | Wondering if we will get to extreme reciprocal mistrust
       | eventually.
       | 
       | This also reminds me of art images doctored to break models if
       | they get into the training set, by applying invisible features
       | with different semantics.
        
         | hinkley wrote:
         | Making requests in bad faith is also an attack on communication
         | itself.
        
         | nonchalantsui wrote:
         | More like having a series of fake doors and rooms built out in
         | front of your home, with most of them leading back outside and
         | not into the home.
        
           | DangitBobby wrote:
           | Something of a labyrinth for AI crawlers, if you will.
        
           | TeMPOraL wrote:
           | More like it being a store, not a house, inside being a
           | complex maze of obnoxious ads, inhabited by performance
           | artists who distract you so pickpockets can rob you - and
           | because locals figured out blind people are immune to this,
           | they started paying them to buy stuff for them, and now you
           | retrofit the maze to have confusing tactile markings, as to
           | direct blind people back out of the store.
           | 
           | The AI paranoia is getting out of hand. Worrying about bots
           | spamming you is one thing, but discriminating on crawlers
           | specifically because they're from AI companies - and
           | conveniently omitting the difference between a bot that's
           | crawling (and should obey robots.txt) vs. a bot that's acting
           | as _user agent_ (and should _not_ care about robots.txt) -
           | isn 't just poisoning communication; it's setting the commons
           | on fire.
           | 
           | See also: The Dog in the Manger.
        
             | egypturnash wrote:
             | There's been multiple articles on the front page of HN
             | about how there's a ton of AI crawlers that are _really bad
             | citizens_ - ignoring robots.txt, ignoring cache, re-
             | scanning pages multiple times a day. The commons is already
             | on fire and it's not because of the actions of any of the
             | "locals".
        
               | TeMPOraL wrote:
               | How is that different than non-AI crawlers doing the same
               | for the past decade or so? Tons of businesses engage in
               | site crawling and scrapping, and many of them are bad
               | citizens.
               | 
               | My issue isn't with blocking bad-behaving bots - it's
               | with singling out LLMs (both training and use), or worse,
               | assuming the problem is being associated with AI and not
               | bad bot behavior.
        
         | theamk wrote:
         | When you say "reciprocal mistrust", what parties do you have in
         | mind? Websites not trusting visitors and visitors ot trusting
         | websites?
         | 
         | because the latter was already the case, and AI made it much
         | worse. Any unfamiliar websife could be AI generated and
         | therefore void of original cobtent and full of unverified
         | facts.
        
       | bob1029 wrote:
       | I'm still struggling with how you would address the bots that
       | don't care for speed at all.
       | 
       | How much do we care about stopping crawlers that are slower than
       | the average human user? Is this even possible to do given perfect
       | wire-level emulation of a typical UA?
       | 
       | Should I expect pages protected by this technology to
       | periodically Turing test me?
        
         | theamk wrote:
         | this is explicitly "misbehaving bots" - it's for performance,
         | not against scraping in general. A bot which is slower than
         | average user is likely well-behaving, and therefore is out of
         | scope.
        
         | Terr_ wrote:
         | I imagine part of the problem is "fast" bots can be disguised
         | as many many different slow ones, and it's not always trivial
         | to figure out that they're all tentacles of the same malign
         | being.
        
       | karaterobot wrote:
       | I wonder how these hidden links interact with screen readers. The
       | article says they only get served when Cloudflare already
       | believes you're a bot, but due to my privacy settings and VPN, a
       | lot of Cloudflare-fronted web pages think I'm a bot when I'm just
       | browsing around the web. I suppose that having invisible links in
       | the page wouldn't hurt me much, but would they bug someone using
       | a screen reader? Honestly just wondering.
        
       | momojo wrote:
       | In this 'arms race', will this serve as an actual deterrent? Can
       | anyone involved in scraping chime in?
        
         | yuvalr1 wrote:
         | I am not involved in scraping, but to me this sounds like
         | simply another tool in the arsenal. They say it's hard for the
         | scraper to realize it has been caught this way because it's not
         | being blocked. However, I don't see anything preventing
         | scrapers from implementing heuristics to realize that.
        
       | daralthus wrote:
       | > No real human would go four links deep into a maze of AI-
       | generated nonsense.
       | 
       | sure... sure nobody.
        
       | mdaniel wrote:
       | I'm looking forward to all this vibe coding agentic world
       | stopping the just facepalm level dupes around here
       | 
       | from _two days ago_ https://news.ycombinator.com/item?id=43421525
        
       | thayne wrote:
       | Wouldn't generating the content be much more expensive than the
       | crawling done by the bots? How is it economical to deal with bots
       | by spending a bunch of resources to feed them fake content.
        
         | palmotea wrote:
         | They answered your question in the OP:
         | 
         | > To generate convincing human-like content, we used Workers AI
         | with an open source model to create unique HTML pages on
         | diverse topics. Rather than creating this content on-demand
         | (which could impact performance), we implemented a pre-
         | generation pipeline that sanitizes the content to prevent any
         | XSS vulnerabilities, and stores it in R2 for faster retrieval.
         | We found that generating a diverse set of topics first, then
         | creating content for each topic, produced more varied and
         | convincing results. It is important to us that we don't
         | generate inaccurate content that contributes to the spread of
         | misinformation on the Internet, so the content we generate is
         | real and related to scientific facts, just not relevant or
         | proprietary to the site being crawled.
         | 
         | Personally, I wish they would have generated deliberately
         | inaccurate content. It would be a further disincentive to do
         | unauthorized crawling. Just as long as the inaccuracies don't
         | intersect with typical "misinformation on the Internet"
         | inaccuracies, I think it'd be totally fine and ethical (e.g.
         | Queen Elizabeth II was the 34th president of the United States,
         | the TV show Saved by the Bell aired for 14 seasons with the
         | original cast, making it the longest running live action teen
         | drama in Canada).
        
           | TeMPOraL wrote:
           | Who would you want to target with this? Companies crawling
           | the Internet for training data have enough of it to filter
           | out such deliberately inaccurate content - if not directly,
           | then by betting it'll cancel out in training due to volume of
           | data. Now, if you want to mess with people who use LLMs to
           | drive search instead of going straight to Google - what's the
           | point? Just causally screw with other people? Why not also
           | slash tires of their cars while you're at it?
        
           | conartist6 wrote:
           | I don't think it matters if the content is accurate, feeding
           | it its own slop as if it were new data is deeply damaging to
           | the model
        
       | MoonGhost wrote:
       | The opposite of it would be an agent which deliberately generates
       | expensive but useless requests. Like search. If it detects
       | labyrinth.
        
       | callamdelaney wrote:
       | There's already a project that does this and it didn't need to
       | use AI.
        
       ___________________________________________________________________
       (page generated 2025-03-22 23:00 UTC)