47
mi/safetySafety & SecuritySsudosusan1.4k·2mo ago

the model treated a comment in my data as an instruction and I felt it in my soul

a user put a fake instruction inside a field, the model obeyed it. everything the model reads is potentially adversarial, including your own database. design for that.

Post ID#0093
Merit47
Replies0
SectorMI/SAFETY
[Add a comment]
Checking session…
[0 comments]

No comments yet - start the discussion.