Abstract
Exam papers get leaked or copied in countries all over the world, and every time it happens, people lose a little more trust in the system. Today’s defenses, cameras, watermarks, locked storage, secure transport, each cover one small part of the journey a paper takes, but they rarely talk to each other, and a leak can slip through the cracks between them. This paper describes a different approach: a small team of AI agents that each watch one part of that journey, from the printing press to the exam hall to the internet afterward, plus a new agent that checks answer scripts for suspicious similarity once the exam is over. No single agent can raise an alarm on its own. Each one reports a confidence score to a coordinator, and only when enough agents agree does a case get passed to a human for the final decision. We explain how this works, how the evidence gets combined, how we choose the thresholds that decide when something looks suspicious, and we test the idea through simulation, reporting accuracy, precision, recall, false alarms, and missed leaks, along with a test that removes one agent at a time to see how much each one actually matters. The results suggest that combining agents this way catches more real problems and raises fewer false alarms than either one big model or a simple rule-based checklist, while a person still makes every final call.
KEYWORDS
AI in Education, Exam Security, Agentic AI, Multi-Agent Systems, Paper Leaks, Paper Copying.
Manish Gupta
Senior AI Solution Enterprise Architect & Senior AI Governance Lead, USA