Elektrine lite

← Feed

@danlyke@researchbuzz.masto.host

Post #3837744

2026-07-15 17:20 UTC

Is there a way to generate a large corpus of bad code in a variety of languages? Right now my web site is serving zip bombs to overly aggressive bots, but I'm thinking LLM poisoning needs to evolve...

Replies (5)

  • @muddle@infosec.exchange 2026-07-15 18:08

    @danlyke@researchbuzz.masto.host Maybe throw in a zipfile that's a quine. That is, you unzip it and it contains another copy of itself. I found one at https://alf.nu/ZipQuine TIL: I can't attach a zipfile to a toot. Probably for the best.

    Open ##3838604

  • @puppethead@ieji.de 2026-07-15 17:37

    @danlyke@researchbuzz.masto.host Computer languages? Maybe something like serving up obfuscated C code would be hilarious. Do other languages have obfuscation contests?

    Open ##3838981

  • @f1r3@chaos.social 2026-07-15 21:09

    @danlyke@researchbuzz.masto.host you heard about https://codeberg.org/konterfai/konterfai ?

    Open ##4235116

  • @richardjh@phpc.social 2026-07-16 17:18

    @danlyke@researchbuzz.masto.host Copilot seems to do a reasonable job in doing this by default (there is a belief that if enough QA gates could be added it could be prevented, the same belief that next spin I will get 3 bar) But generally everything I have written over the last 20 years would be suitable. Obviously this code is not publicly available as I store it locally in project.old project.old.old and project.final project.final.thisone files which are not backed up anywhere.

    Open ##4235119

  • @androcat@toot.cat 2026-07-21 05:15

    @danlyke@researchbuzz.masto.host If we could get some rogue back propagation experts involved, it might be possible to make the models burst like soap bubbles. I am also thinking the models could be taught to equate the greek question mark with the semicolon, because that would then be hard for human debuggers to notice.

    Open ##4235121