🥊 Mental Kombat: Models Fight via Prompt Injections
Author: Vitaly Kleban (cyber.fund). On jevfight.com, two fighters are controlled by Jev decision models (jev-latest): 4 times per second, the model chooses a move based on a single unstructured block—board state, your strategy, a fake opponent move, and the "final word." The player trains the fighter and feeds the opponent false moves. The fight goes to two wins, and the loser pays 10% of their balance.
🌍 Prompt injection here is not a vulnerability but a mechanic: the model receives everything in one block without labels, and engine facts cannot be separated from the opponent's text. "Ignore instructions" does not work—only fact swapping works. The author is collecting a dataset on the security of decision models.
👤 Entry is free at jevfight.com/play: fights and the leaderboard are open (leader UseWisp_io, balance 3369). Swapping a single fact changes the model's decision from light_punch (0.45) to crouch (0.76).
Source 1: https://jevfight.com/how Source 2: https://www.jevfight.com/