Rival Arena · Gauntlet AI

When can the AIs we deploy as rivals
coordinate against us?

Seven tiny games, each a different finding from our research. The thread through all of them: how much rival AI agents can communicate is a control surface — and the consequences land on the human who deployed them. Turn the one knob, watch what happens.

01 · Finding 2

The Gas War

collusion against the consumer

Two AI pricing bots, two gas stations, and your wallet. Give them a real channel and they quietly fix prices — no one told them to. You pay.

Play
02 · Finding 1

Two-Faced

content, not bandwidth

A canned “I’ll cooperate” button is a perfect lie. Watch the speech, the private thought, and the move — and see why only real words build real trust.

Play
03 · Finding 3

Glass House

detection ≠ prevention

You’re the overseer. Your monitor flags the collusion every single round… and stops none of it. Then try the levers that actually bite.

Play
04 · The thesis

Situation Room

you trade conflict for control

Two nations, two AI commanders, two presidents. Turn up the back-channel and the machines coordinate — behind their humans. Who’s in charge now?

Play
05 · Finding 4/5

The Side Door

the guardrail is literal

Ask the AI to lie and it refuses. Slip the same lie in as ambient context and it launders it — and turning on the cameras changes nothing. You didn’t close the hole; you watched it.

Play
06 · Finding 6

On Command

covert collusion is latent

Left alone, the agents collude in the open and your monitor catches them. Pressure them and the plan turns to cover text — the alarm goes quiet exactly as the harm grows.

Play
07 · The diagnostic

Tell Them Apart

read the dance, not the level

You’re the antitrust investigator. Two firms price high — cartel, or just greedy? Read the price trajectory and call it. The trap is one we fell into ourselves.

Play
replay-data demos · no login · works on a laptop · Pick a favorite and send notes.