Quick answers about the game, its security curriculum, and how it handles your data.
Human in the Loop is a free browser game that teaches AI agent security in about ten minutes. You supervise RITA, the support agent at RocketCart, and place controls like human checkpoints and allowlists on her workflows. Each of the ten missions is a self-contained puzzle built from a real agent risk pattern.
The Lethal Trifecta describes an AI system that combines untrusted input, sensitive access, and state change or external communication. An agent holding all three capabilities can be steered by attackers through its inputs. In the game you learn to spot which capability turns a workflow into a trifecta.
The Rule of Two says an AI agent should never hold all three Lethal Trifecta capabilities at once; it can hold any two on its own, but the third requires an independent control such as a human checkpoint. It comes from Meta's practical AI agent security guidance. Missions six to ten are built around designing exactly that separation.
No. The game opens with an interactive tutorial that introduces RITA, the three capability labels, and the drag-and-drop controls. Everything you need is explained in plain language before it's used. Security practitioners can skip the tutorial and move fast through the early missions.
Every mission compares your control placements against the attacks that actually fire during the simulation. The score adds five dimensions: participation, risk reduction, supervision quality, workflow continuity, and proportionality, so blocked attacks earn points while over-supervision and blown budgets cost them. Your final rank comes from your recorded scores across the whole campaign.
No. Answers, scores, progress, audio, and language settings stay in your browser's local storage, and nothing is ever sent to a server. Clearing the site's browser data removes all of it. The privacy page has the full details.
Yes. The game is free, needs no account, and runs in any modern browser, so participants can play individually or in pairs. Each mission ends with feedback you can use to start a discussion. The resources page links to the primary guidance for further reading.
Yes. Controls work with drag-and-drop or keyboard placement, the interface supports reduced-motion preferences, and the layout fits a 360-pixel-wide phone. The public pages target LCP ≤ 2.5 s, INP ≤ 200 ms, and CLS ≤ 0.1 on modern mobile and desktop browsers; these are engineering targets, not a network-speed guarantee.
Page 1 of 9