Why swap an inference provider's model for an abliterated one? "Refusals can be REALLY annoying"

moyix · x · 2026-11-20

Cryptographer Matthew Green asked why anyone would compromise an inference provider and swap in an abliterated (censorship-stripped) model. AI security researcher moyix offered a tongue-in-cheek answer: "Look, refusals can be REALLY annoying," poking fun at the friction of model safety guardrails.

Related event: Cryptographer Asks Why a Hacked Inference Provider Was Swapped for an Uncensored Model(3 posts)→

Original post →

More from Fun

Fun channel →