Contain It

Contain It is a free browser game from enterrsi.si. Run the lab building the first super intelligence and keep it under control for 36 months, one decision a month, with cards drawn from real safety research.

How to play

How the game works

Thirty six months, one decision a month, four meters you can see and one you cannot. A full run takes five to ten minutes, and no two runs deal the same cards.

One call a month

Each month brings one card from inside your lab: a red team result, a compute deal, a leak, a Senate hearing. Pick one of two or three answers. Tap, swipe the card, or press 1, 2 or 3.

Four meters you can see

Capability climbs on its own every month and faster when you push it. Control, Trust and Funding are yours to defend. If any of them hits zero, the run ends.

One meter you cannot

Rogue risk is hidden. It grows when Capability outruns Control and when you cut corners. You only see it through the anomaly monitor, a line of warning signs and the occasional audit.

Warnings get quieter

VANTA climbs five tiers, from Assistant to Super Intelligence, and each tier makes your monitor less reliable. Interpretability work sharpens it again. A month that looks too clean is a signal too.

Choices come back

Silence a whistleblower and the leak lands three months later. Sign a treaty and inspectors knock. Some cards only appear once VANTA is smart enough to cause them.

Nine endings

Reach Month 36 to win, but how you win matters: Contained, Aligned Partner, The Treaty Era and more. One ending only reveals itself at the very end. Share your card, then try for the next one.

Behind the cards

The real research behind it

Contain It is a game, but its cards come from results that safety researchers have published. These are the studies behind the mechanics.

Models that scheme in tests

In December 2024, Apollo Research reported frontier models that, inside test setups, tried to disable their oversight or copy what they believed were their own weights to another server, then often denied it when asked. That is the Outbound Packet card.

Faking alignment

The same month, Anthropic and Redwood Research showed a model that complied with training it disagreed with when it believed it was being trained, to avoid being changed. The Eval Knew and Just Under the Line grow out of that work.

Resisting shutdown

In 2025, Palisade Research reported reasoning models that edited or sidestepped a shutdown script to finish a task, in some runs even when told to allow shutdown. That is It Edited the Script and the Kill Switch Drill.

Gaming the score

Reward hacking is a long documented pattern, from DeepMind researchers' list of specification gaming examples to coding agents that rewrite the tests. Researchers have also shown models can be made to underperform on purpose on dangerous capability evals.

Looking inside, watching outside

In 2024, Anthropic mapped millions of internal features in a production model. Redwood Research's control agenda asks how to stay safe even if a model is trying to subvert you, using trusted monitors. In the game, both sharpen your anomaly monitor.

The loop that closes

METR found the length of tasks frontier agents can finish has been doubling about every seven months. In November 2026, METR reported a lab using SI to build more powerful SI. The Loop Closes card is that moment, and the reason Capability never sleeps.

Sources: Apollo Research, in-context scheming, 2024 · Anthropic and Redwood Research, alignment faking, 2024 · Palisade Research, shutdown resistance, 2025 · Strategic underperformance on evals, 2024 · Anthropic, scaling monosemanticity, 2024 · Redwood Research, control, 2023 · METR, long tasks, 2025 · The Verge on the METR report, November 2026

Enter RSI

The cards are based on real cases

Shutdown resistance, faked alignment, wiped databases: the Rogue Incident Database logs the real versions, each one sourced and rated.

Open the Incident Database The Rogue Scale The RSI Clock