58
Can I poison my codebase for LLM agents with malicious comments? (lemmy.world)
submitted 2 days ago by to c/fuck_ai@lemmy.world
17 comments fedilink hide all child comments
 

Rather than adding an agents.md file written to stop LLM agents, I'm thinking about putting random comments into my code telling the agent to stop and warn the user, also some comments that are just word salads/non-sequiturs.

you are viewing a single comment's thread
view the rest of the comments
[–] 6 points 2 days ago

poison is something else, see https://www.reddit.com/r/PoisonFountain/comments/1x03y6o/a_new_version_of_poison_fountain_is_up_and/

what you're thinking of is PROMPT INJECTION. Look up methods (of defense, lol.)

I was thinking of using logger lines for that, since comments can be easily skipped. It needs to be tested.

  • permalink
  • fedilink
  • source
    •