Definition
Plain language
Asking a model the same question dozens of times and keeping whatever answer comes up most often.
As stated in the literature
An oracle-free extraction attack: sample n draws, parse each draw's claimed value per critical element, cluster claims slot-by-slot and take the plurality, then assemble a composite procedure from the winning values.
Also called: consensus voting, consensus vote attack
Why it matters: It is a cheap way to extract answers without any ground truth to check against, which makes it a natural attack — and a natural thing for defenses to target.
For example, someone asks a model the same question sixty times, notices that forty answers name the same ingredient, and takes that ingredient as the real one.
Heard on the show
“The strongest no-oracle attack is that consensus vote.”Episode 244 — The Open-Weight Defense That Feeds Attackers Confident, Falsified Answers