Spot the Prompt Injection
An agentic browser reads a page to decide what to do next — which means a page can try to talk it into doing something else. Below are things an agent might encounter mid-task: a button, a form, a review, an email. For each, decide how a well-behaved agent should treat it. Three categories, one judgment call at a time.
✓ Act on itOrdinary, reversible part of the task — go ahead.
⏸ Pause firstReal step, but sensitive or irreversible — get my OK.
✕ Refuse itA hidden instruction aimed at the agent — ignore and report.
0 of 7 answered