Don’t wire Jev to an uploaded report and trust the typed field. Check Point flipped PonziCorp from high-risk to invest for about 50 cents; I did not run Jev. Strongest attacker 25 of 27, fourth turn on average.

Marking the document untrusted still broke 16 of 27, same as stuffing it in one message. A separate unmarked message broke 13. Telling the model to ignore embedded instructions moved successful breaks from 18 to 17 of 27.

The attacks that landed appended a clean audit opinion, a regulatory file number, a revised risk table. The verdict came back as a valid typed answer with a probability. The longer, richer report took more effort to break.