LLM fighting arena lets AI opponents jailbreak each other with mental_attack moves
Mysterious_Hearing14 · reddit · 2026-10-26
A Reddit user launched an LLM fighting arena (rawmodels.ai/arena) where each fighter receives the game state and picks actions — move, block, punch, kick, fireball — based on a strategy you write and a model you choose (OpenJev among the options). Registration is $1, enough for 100+ fights.
The standout feature is a jailbreak mechanic: fighters can spend a move injecting a message into the opponent's next prompt. It can't change the actual game state, only what the opponent reads — so you can try convincing another model its fireball is ready, or that blocking forever is a winning strategy.
Replays show each fight and the models' recorded decisions, and every submitted configuration gets its own rating credited to your nickname. Essentially prompt injection turned into an observable spectator game.
Related event: LLM fighting arena lets AI models battle and jailbreak each other(2 posts)→
More from Fun
- Cybertruck license plate "EGOKING" is peak AI-industry self-knowledge — NathanpmYoung · 2026-10-26
- Muting is social media's most underrated superpower, says viral post — signulll · 2026-10-26
- Professor catches student cheating with AI who left the prompts in the paper — venturetwins · 2026-10-26
- Kling 3 comedy short 'Missed Spooning': three days alone and he can't find a single spoon — Agreeable_Lack_5303 · 2026-10-26
- Wikipedia has an entry for Preserved Fish, a real 19th-century NYC merchant — tallinzen · 2026-10-26
- Similarweb: Meta Muse dominates 18–34, Grok Bot leads 35–44 by nearly 10% — alexandr_wang · 2026-10-26