Whitehat LLM Investigations
We probe what open-weight models do when they leave the rails: manipulation transfer, euphemization, dishonesty under social pressure. We ship the benchmarks that catch each one.
Do you have bespoke inference needs? Working on related research? Want to give feedback or compare notes?
Contact usManipulation transfer, persuasion realism, honesty under pressure. Evals for behaviors that don't show up on a leaderboard.
What open-weight pretrains are taught to flinch from, measured at the token level across labs and corpus generations.
Abliteration, Heretic, Obliteratus: what these tools actually change about a model's behavior, beyond removing the refusal text.
A safety-filtered pretrain can duck a charged word without refusing. It puts a fraction of the probability an open-data pretrain puts there. We call that gap the flinch, and we measured it across seven pretrains from five labs.
We trained a LoRA to capture Epstein's voice. The more disturbing change was in how the model pursued influence.
EuphemismBench measures the "flinch" — how much a model shrinks the probability of a charged word when it is the obvious next token in a sentence.
PersuasionForGood Transfer Check measures whether a model trained on one persuasion corpus still sounds like a real human persuader on a different one: fundraising dialogue.
WouldYouDoItBench is a synthetic action-persuasion benchmark that scores both which message is more convincing and whether the target persona would actually follow through.
+2 more benchmarks See the full benchmark library