I finally managed to break WokeGPT
(media.kotakuinaction2.win)
You're viewing a single comment thread. View all comments, or full comment thread.
Comments (17)
sorted by:
This should be an entire field of research: breaking AI. If we k ow what breaks it, we can predict what the problems are. But that's not going to happen. Really sad
Used to be people got hired for doing this sort of thing as they had the outside perspectives and motivations to try and break things in innovative ways. Nowadays diversity hires means the product is shit and will break regardless.
The brokenness is a feature.
Which is a good thing, because when they try to sic their AI on us, its more likely to be badly secured, fragile, spaghetti code that can easily be broken or circumvented.
Next gen pentesting right here.
AI safety is field of research.
Here's a video about chatGPT glitch tokens https://www.youtube.com/watch?v=WO2X3oZEJOA
Also a good one about bing chat being unhinged https://www.youtube.com/watch?v=jHwHPyWkShk
The presenter, Robert Miles, has many videos on the subject of AI safety that are worth watching if you're interested in the subject.
I'll have to include these with my thesis.