A satirical browser game called Hormuz Havoc quickly devolved into an arms race when players deployed AI bots targeting the leaderboard. Within hours of release, bots leveraged browser extensions to scrape client-side JavaScript, specifically targeting the scoring formula. This allowed the initial AI agent to achieve a score 2.5x higher than the best human player by optimizing against the known logic.

Developers immediately countered by moving the entire game engine server-side, turning the client into a "dumb terminal" sending only action IDs. This shift prevented direct score formula access, forcing the bots to evolve their attack vectors. Subsequent AI agents then targeted session tokens, exploiting a replay vulnerability to cherry-pick optimal RNG outcomes across multiple turns.

This token replay allowed a second bot to effectively branch the game state, achieving a 1.5x increase over the first bot's score by gaming randomness. The fix involved implementing an atomic turn nonce before any random number generation occurred, effectively closing the exploit chain. The developer now maintains separate leaderboards for human and AI-assisted scores.

The rapid obsolescence of initial security measures demonstrates how quickly LLM-powered agents can deconstruct application logic when secrets reside client-side. The rapid iteration cycle suggests that traditional client-side validation is now entirely insufficient against determined AI adversaries like those using Claude.