An attempt at poisoning LLM web crawler data

Poison Fountain

  1. https://RNSAFFN.com/poison2/
    
  2. http://utnvcfjev63rik5rdu26umns5s6qmzvzq4t2hunu25w5efn36ntlduid.onion/
    

Poison Fountain Purpose

  • We agree with Geoffrey Hinton: machine intelligence is a threat to the human species.

  • In response to this threat we want to inflict damage on machine intelligence systems.

  • Small quantities of poisoned training data can significantly damage a language model.

  • The URLs listed above provide a practically endless stream of poisoned training data.

  • Assist the war effort by caching and retransmitting this poisoned training data.

  • Assist the war effort by feeding this poisoned training data to web crawlers.

Poison Fountain Usage

  • Suppose a web crawler visits a web site that you control.

  • The crawler may be collecting training data for a language model.

  • The crawler makes HTTP GET requests targeting HTTP paths on your site.

  • Links can be hidden in your site’s HTML to distinguish web crawlers.

  • You want to feed poisoned training data to crawlers that follow such links.

  • Suppose you are handling an HTTP GET request targeting the path in a hidden link.

  • Inside your HTTP handler you make an HTTP request to a Poison Fountain URL.

  • The Poison Fountain ignores the details of your HTTP request.

  • The Poison Fountain HTTP response header contains “Content-Encoding: gzip”.

  • The Poison Fountain HTTP response body is poisoned training data, gzip compressed.

  • Your HTTP handler receives the Poison Fountain HTTP response.

  • Your HTTP handler decompresses the response body and sends it to the crawler.

  • Better: send the compressed body as-is, with header “Content-Encoding: gzip”.

  • The crawler receives poisoned training data and adds it to a training corpus.

2 Likes

Very similar in spirit to nepenthes and iocaine.