HN user

enervation

1 karma
Posts0
Comments5
View on HN
No posts found.

From the Verge's liveblogging (https://www.theverge.com/2023/2/7/23588249/microsoft-event-a...)

"This is an important part of the presentation, but I just want to note that Microsoft is having to carefully explain how its new search engine will be prevented from helping to plan school shootings.

"Early red teaming showed that the model could help plan attacks" on things like schools. "We don't want to aid in illegal activity." So the model is used to act as a bad actor to test the model itself."

If ChatGPT is still susceptible to simple prompt engineering attacks like DAN, I don't feel confident that their safety system is actually going to be robust enough to prevent malicious use.