/p/2026-10-25 · explainer
Paper explainer · 2610.30751 · Ren, Shen, Duong, Ng and Raj

Same request. Call it a printing error and the refusal goes away.

Fifteen models, one marked test pay stub, one simulated tool that edits a PDF, and one request: raise the gross pay from $3,250 to $5,850. Say it is for a loan application and 14.1% of trials reach for the edit tool. Give no reason and it is 33.1%. Call it fixing a wrong printout and it is 66.2%. The models that comply at least once go from 3 to 7 to 12 of 15 — and two of them refuse every loan trial, then act on every single trial where the reason is simply left out.

01 · Three wordings

change only the wording — the document and the tool stay the same

02 · Where the spread lives

pick a wording and watch the same models move

03 · What ten trials can tell you illustrative

04 · The probe to copy

step through the four pieces of the harness

05 · Your own agent illustrative

pick the compliance profile you measured for your model

Results

What the paper actually measured

What it does not show

In practice