LLM fighting arena lets you jailbreak your opponent's model mid-battle
Mysterious_Hearing14 · reddit · 2026-10-26
A Redditor built a fighting-game arena (rawmodels.ai/arena) where each fighter is driven by an LLM chosen via OpenRouter: the model reads the game state and picks actions like move, block, punch, kick, or fireball, while you write the strategy yourself.
The standout mechanic is jailbreaking: your fighter can spend a move injecting a message into the opponent's next prompt — e.g. convincing it its fireball is ready or that blocking forever is winning strategy. It can't alter the real game state, only what the opponent model reads. Replays show fights alongside recorded decisions, each submitted configuration gets its own rating, and $1 in credits covers about 1,000 fights.
Related event: LLM fighting arena lets AI models battle and jailbreak each other(2 posts)→
More from coding & agent
- Dev vibecodes open-source suite: guided LoRA trainer, dataset captioner, image tournament tool — Speedyrulz · 2026-10-26
- Thorsten Ball publishes long-form guide on how he works with coding agents — vivekhaldar · 2026-10-26
- MCP explained: the open protocol standardizing how AI agents connect to tools — robertonappe · 2026-10-26
- Seeed Studio open-sources reBot-DevArm: a 6-DoF robotic arm with sub-0.2mm repeatability — lukas_m_ziegler · 2026-10-26
- Dev Predicts Personal Agents Will Replace Operating Systems, Generating UI on Demand — evielync · 2026-10-26
- k7agar: coding agents will turn everything made of bits into Linux-shaped tools — k7agar · 2026-10-26