Poisoning Well for LLMs

Re:Poisoning Well for LLMs by kitku on Friday September 05, 2025 @04:15AM (#45136207)

This reminds me of the Nepenthes tarpit [1], which is an endless source of ad-hoc generated garbled mess which links to itself over and over.

Probably more effective at poisoning the dataset if one has the resources to run it.

[1]: https://zadzmo.org/code/nepenthes/

[ Parent | on HN ]

Re:Poisoning Well for LLMs by fleebee on Friday September 05, 2025 @05:26AM (#45136630)

I'm running Iocaine[1] which is essentially the same thing on my tiny $3/mo VPS and it's handling crawlers bombarding the honeypot with ~12 requests per second just fine. It's using about 30 MB of RAM.

[1]: https://iocaine.madhouse-project.org/

[ Parent | on HN ]

Re:Poisoning Well for LLMs by treetalker on Friday September 05, 2025 @07:12AM (#45137265)

Odorless, tasteless, and among the more deadly poisons known to crawlers!

[ Parent | on HN ]

Re:Poisoning Well for LLMs by BrenBarn on Friday September 05, 2025 @04:09PM (#45143048)

Unfortunately they will spend the next several years building up an immunity.

[ Parent | on HN ]

Re:Poisoning Well for LLMs by 8organicbits on Friday September 05, 2025 @06:00PM (#45144161)

Do we know if LLM scrapers are running JavaScript on the pages? If they are, maybe it's worth offloading the Markov model to the client side.

[ Parent | on HN ]