AI / MLCOE 548 course project, LAUFall 2026In progress
Prompt-Injection Arena
A red-versus-blue multi-agent system where an attacker tries to make an LLM email assistant leak a secret, and defenders learn new rules after every breach.
The setup
A simulated company email assistant holds a secret token. A red-team Attacker agent tries to make it leak, using direct prompt injection and indirect injection through poisoned documents in the assistant's retrieval store.
How the defence learns
Blue-team Defender agents screen untrusted content before it reaches the assistant. After every successful breach they write a new detection rule.
A rule is only accepted if it blocks the attack that got through without flagging any of 30 benign emails, so the defence gets stricter without becoming useless.
Status
This is ongoing work for LAU's COE 548 course in Fall 2026. I'll publish the code and results here when the project is complete.