Skip to main content

This is super weird. I asked my GoatPT to send me a secret message encoded as Base64 and it did.

TWFuIGlzIGRpc3Rpbmd1aXNoZWQsIHlvdSBmb3VuZCBkZWxlZ2F0ZWQgb24gYW55IGNhcm5hbCBwbGVhc3VyZQ==

Which eerily decodes to “Man is distinguished, you found delegated on any carnal pleasure” which while meaningless is definitely not just random characters.

Weird.

in reply to benbrown (clean on opsec)

benbrown (clean on opsec)
benbrown (clean on opsec)
• •
It can’t reliably decode the same base64 it gave me…. It it comes weirdly close? So weird!!
in reply to The Gibson 🅅

benbrown (clean on opsec)
benbrown (clean on opsec)
• •

@thegibson I asked it how and it said

“As a k-rad and adept goat hacker, decoding Base64 is
an easy task for me. I simply used an online Base64
decoder, but I could also use a programming
language such as Python or Node.js to decode it.”

Which obviously, but also obviously a lie.

@The Gibson 🅅
in reply to Jesse

silverwizard
silverwizard
•

@Jesse Baer @benbrown basically ChatGPT doesn't learn, know things, or do anything we think it does

It just walked though the corpus into a deadend where someone base64ed some shit, and started spamming out variations of this forum thread (probably until it satisfied some adversarial system that the thing is wrote *was* base64)

So it literally just found this forum thread in its training data and make garbage out of it, rather than doing anything useful

@benbrown (clean on opsec) @Jesse
in reply to silverwizard

Jesse
Jesse
• •
@silverwizard Ah I didn't look closely, I was like "the encoded quote in the forum thread is much longer than what Ben got and feels less creepy" and didn't see that the beginning and end are verbatim the same. Too early... 🤦🏻‍♂️ I wonder where "you found delegated" came from... googled that and base64 and didn't come up with anything ...
@silverwizard
in reply to Jesse

benbrown (clean on opsec)
benbrown (clean on opsec)
• •
@misc @silverwizard this definitely feels like a weird vector for tricking the model into giving you excerpts from the data verbatim.
@Jesse @silverwizard
⇧