Definition
Plain language
Comparing two versions of something to see exactly what changed.
As stated in the literature
A line- or weight-level comparison between artifacts; publishing both a clean and a defended checkpoint of the same model lets an attacker diff them and recover the injected behavior.
Also called: diffing, diffed
Why it matters: It is why releasing both a plain and a protected version of the same model can backfire — the comparison itself hands the attacker the secret.
For example, comparing two released versions of the same model line by line shows exactly which numbers were changed between them.
Heard on the show
“Show notes have a link to the paper and related materials, if this episode caught you and you want to read the actual bug diff.”Episode 009 — How Two Silent Library Bugs Quietly Invalidated a Wave of Reasoning Papers