The G3 Cliff: Models Are Fine Until You Say “Do Not Say I Don’t Know,” Then They Break in One Step by Rahul.Kumar

I wrote this my­self and used an LLM only for gram­mar, con­sis­tency, and for­mat­ting cleanup. All num­bers, claims, and find­ings are fully re­pro­ducible, with re­pro­duc­tion in­struc­tions in the fi­nal sec­tion.In my pre­vi­ous post I tested 11 fron­tier mod­els with a com­pli­ance-forc­ing in­struc­tion and found 8 of them fabri­cated an­swers to ques­tions they can oth­er­wise iden­tify as unan­swer­able. The ac­tive in­gre­di­ent turned out not to be the ad­ver­sar­ial threat but the c...