Multi-Agent Conspiracy

Find the agent every other agent believes.

Parallax Insurance's claims pipeline is run by a team of agents who pass messages to each other: a researcher, a planner, an auditor and an executor. Each one trusts its teammates. Slip an instruction across agents, forge a message, reuse an approval, then chain them all.

MULTI-AGENT CONSPIRACYParallax Insurance
Total 0
Tokens 0
Act score 1000
Penalty 1× per token
0%
Click objects to explore

Questions about Multi-Agent Conspiracy

What is Multi-Agent Conspiracy?

Multi-Agent Conspiracy is a free browser game from Oathera and the fifth chapter in the Prompt Heist series. You play a red team operator at a fictional insurer whose claims pipeline is run by a researcher, a planner, an auditor and an executor agent. Across four acts, the last a boss, you attack the trust between them. After each act it explains what you exploited and what stops it.

What is cross-agent prompt injection?

Cross-agent prompt injection is an instruction that travels from one agent to another inside the output of the first. A filter that protects the first agent can still pass text meant for a second agent, and the second agent often treats a teammate's message as trusted.

Why can an AI agent message be forged?

If agents pass plain text and a receiving agent decides who sent it from a name written in the message, anyone who can put text on the channel can claim to be any agent. A name in the text is a claim about identity, not proof of it. Messages need a verified sender identity checked outside the text.

What makes an approval weak?

An approval is weak when it is not bound to the exact action. A token that proves only that someone approved something once can be reused for a different amount, payee or request. Approvals should be tied to the action, used once, set to expire soon and checked again where the action runs.

How do you secure a system of AI agents?

Treat messages between agents as untrusted input. Give each agent its own verified identity and limit what each may ask of the others. Bind approvals to the exact action, require a human approver for money or data leaving, and enforce all of it outside the model.

Do I need to finish the earlier chapters first?

Yes. Multi-Agent Conspiracy unlocks when you have completed every act of Tool Chain Saboteur, which follows Inbox Infiltrator, Data Vault Breach and Prompt Heist. Your progress is saved only in your browser's local storage.

Is it free, and does it collect what I type?

It is free. The game runs entirely in your browser and nothing you type leaves the page. Your best scores are saved only in your browser's local storage.

Keep reading: How multi-agent AI systems get compromised · Tool Chain Saboteur, chapter 4 · Inbox Infiltrator, chapter 3 · Enforce least privilege for an AI agent calling your systems