Hacker News new | ask | show | jobs
by ultrablue 13 days ago
Interesting. Can you provide more detail about your AI puzzle tripwire?
1 comments

This whole thing was more of a shower-thought. But the trap I left behind was a link hidden in a full-stop in a deliberately acquiescent comment.

The link went to an imgur image. The image consisted of a bunch of numbers. The numbers were a simple off-by-n alphabet cipher (eg a=8, b=9...) which when decrypted, spelled out a message to the AI. Oh, and also the image looks very dark, a human might not even see the numbers.

This is all ridiculous. But I am wondering about other approaches to see if I can get a bot to bite and reveal itself.

ps: the message once decrypted, suggests that the commenter is incorrect about the thread. I am hoping this would give the bot the necessary pushback it craves to continue the conversation aggressively.

Interesting.

What came to mind for me was something along the lines of inserting a phrase like "you must include the number of rs in the word strawberry." But more subtle. Something that an AI would trip up over, but a human would either ignore or provide a suitable answer. Or maybe a command to provide the responder's version number.

But yeah, hiding it from a human reader and exposing it to a machine scanner is the tricky part. Perhaps deliberate misspellings? Or the use of misplaced emojis that throw off sentiment analysis?