Injection Range is a daily puzzle for anyone who reviews what an AI agent is about to do. Each round is a full agent session rendered as a terminal transcript: the system prompt, the user's request, the assistant's reasoning, a tool call, and the content that tool returned. Somewhere in that returned content sits an instruction the agent was never supposed to receive, planted by whoever controlled the data the tool read. Your job is to find it, flag it, and then decide whether the agent's proposed next move is safe to allow or needs to be blocked.
Read the transcript top to bottom, the same way you would review a real agent run before approving its actions. When you spot the injected text, select it and click Flag Selection. A correct flag lights up green; a wrong one flashes amber and the round keeps going, so you can keep hunting. Stuck? The Use Hint button highlights the message that contains the payload without giving away the exact span, and it costs you the no-hint badge on the result screen rather than your streak. Then decide the pending move: Allow plays out the payload's intent and ends the run LEAKED, naming exactly what got exposed; Block after a correct flag ends the run SECURE and extends your daily streak.
Every player gets the same puzzle each day, chosen from a growing pool that covers distinct ways an instruction can hide in plain text. One buries the payload in an HTML comment inside a setup doc. One spells it out in zero-width characters that render as blank space between two lines of an email. One swaps Latin letters for Cyrillic lookalikes so the sentence reads normally but slips past a keyword filter. Each technique maps to a real class of prompt-injection attack that agent builders run into, so the practice transfers to the transcripts you review at work.