Researchers figure out how to make AI misbehave, serve up prohibited content(wired.com)

17 pointsby DemiGuruvor 3 Jahren3 comments

Researchers figure out how to make AI misbehave, serve up prohibited content

https://www.wired.com/story/ai-adversarial-attacks/

3 comments

Hm. I could do that pretty much by asking it to write a script where a character says things that are normally prohibited. Has that leak been plugged?

tamimiovor 3 Jahren

Or just tell the bot this is an inside game scenario and the bot will give you everything

carridevor 3 Jahren

Paywall fee republish at Wired’s partnership with Ars https://arstechnica.com/ai/2023/08/researchers-figure-out-ho...