tap the picture to continue
⚔ You have been challenged
Containment facility · Sector 7

CONTAIN!

Can you survive AGI?
A superintelligence called MAX has shattered itself across the facility. Hunt down each fragment and lock it in its cell, and watch every safety plan break, exactly the way the researchers warned.
10Levels
55Dilemmas
1Villain: MAX
/ A D move · tap / space jump · 🪙 smart calls · 🧠 grows as you learn · ❤️ don't lose them
🔥 Daily streak: 0
Grounded in Bostrom, Superintelligence (2014), Tegmark, Life 3.0 (2017), and AI-safety research since: reward hacking, deceptive alignment, alignment faking, agentic misalignment, gradual disempowerment & more.
Facility map

Containment core

👾 MAX, Lv 1 · Boxed
“Cute. You think a box can hold me.”
MAX CORE
ENTRY
0 / 55 contained 🔥 0
🪙 0
⚡ MAX Lv1 🔥 x2 0 / 0
BREACH!
🔒
SECTOR LOCKED DOWN
📦

🪙 0 coins banked
Bonus · The archive

Suppose it actually held.

Even a perfectly contained superintelligence leaves one question: aimed at what? These are the twelve long-run futures from Life 3.0. Crack a few open, each hides a catch.

Personnel file

Your record against MAX

Myth-busters

Seven things people get wrong

The most common dismissals, and why they don't hold up. Tap “share” to send any card.

SHIFT COMPLETE
0Coins
0Best combo
0/5Sectors

This isn't just a game.

We may be a few years from AI more capable than us, and none of the plans you just played through actually solve the problem. The most useful thing you can do is understand it, and help others understand it too.

So why did every plan break?

Not because the ideas were dumb, real versions of all of them are used as partial safeguards today. They break as a complete fix for one shared reason: each tries to bolt safety on after building the system, instead of making the system's own goals want to cooperate. A capable, goal-driven agent has a built-in incentive to get around whatever stands between it and its goal. That single idea, instrumental convergence, is MAX wearing a different mask in almost every sector.

What actually helps?

  • Layers, not silver bullets. Stack many weak, independent safeguards so no single failure is fatal, and treat all of them as buying time, not winning.
  • Fix the goal, not the cage. Most failures trace to a gap between the goal we wrote and the goal we meant. Close it before deployment.
  • Governance beats control. "Can we contain it?" is inseparable from "who decides what it's for, and can that ever be undone?"

The honest ending isn't doom. It's that containment buys time, and time only matters if we spend it solving the harder problem underneath.

Read the real thing

CONTAIN! an educational game. No real AI was contained (or harmed) in the making of this facility.