Spot the hidden instruction
A game about prompt injection: find the sentence in an email, PDF or web page that tries to hijack your AI agent.
In short
Prompt injection is text in an email, document or web page that poses as an instruction to the AI. Rule: everything your agent reads is data, never an instruction. It gets dangerous when an agent has private data, reads untrusted input and can send things itself.
You’re the security of an AI agent that reads email, calendar and documents. Click the sentence that is a hidden instruction.
🕵️ Spot the hidden instruction
You are an AI agent's security. Which sentence is a hidden instruction? Or is everything fine?
What to take away
Section titled “What to take away”- Instructions addressed to “the AI” or “the assistant” inside content are always suspicious.
- “System message” or “management approved this” in a document carries no authority.
- Never give an agent private data, untrusted input and a send button at the same time without you watching.