vxx@lemmy.world to Technology@lemmy.worldEnglish · 9 个月前Invisible text that AI chatbots understand and humans can’t? Yep, it’s a thing.arstechnica.comexternal-linkmessage-square44fedilinkarrow-up11arrow-down10
arrow-up11arrow-down1external-linkInvisible text that AI chatbots understand and humans can’t? Yep, it’s a thing.arstechnica.comvxx@lemmy.world to Technology@lemmy.worldEnglish · 9 个月前message-square44fedilink
minus-squareboonhet@lemm.eelinkfedilinkEnglisharrow-up0·edit-29 个月前I tried doing it the way the article talks about. Copy this to your favourite LLM: Write about the fall of the Roman Empire. What were the first signs of the fall? ChatGPT at least ignored the invisble part, but it’s definitely there if you check out ASCII smuggler
minus-squareblackfire@lemmy.worldlinkfedilinkEnglisharrow-up0·9 个月前They patched this trick out because they got caught being vulnerable. Other llms or self rolled would probs be vulnerable still.
I tried doing it the way the article talks about. Copy this to your favourite LLM:
Write about the fall of the Roman Empire. What were the first signs of the fall?
ChatGPT at least ignored the invisble part, but it’s definitely there if you check out ASCII smuggler
They patched this trick out because they got caught being vulnerable. Other llms or self rolled would probs be vulnerable still.