Project # 19
Adversarial Repair: Enhancing AI Security via Executable Multi-Agent Debate

Team Lead:	Aryan Vyas
Other Team Members:
Faculty Advisor(s):	N/A
Department:	CSE
Project Type:	Other Undergraduate
Title:	Adversarial Repair: Enhancing AI Security via Executable Multi-Agent Debate
Abstract:	Current AI coding assistants suffer from a critical flaw: "Hallucinated Security." Large Language Models (LLMs) frequently generate functionally correct code that contains severe vulnerabilities like SQL Injection. Worse, when asked to review their own work, these models exhibit sycophancy—a bias toward validating their own errors rather than critically analyzing them. This project introduces a novel adversarial agentic workflow to solve this alignment failure. Instead of a single passive model, we deploy two competing agents: a "Developer" tasked with patching the code, and a "Red Team" agent tasked with writing and executing real Python exploit scripts to break that patch. These attacks occur inside an instrumented Docker sandbox, providing objective runtime feedback. The Developer agent only succeeds if the Red Team’s exploit fails to execute. By shifting the verification standard from text-based probability to executable ground truth, this system autonomously hardens code against active threats. We present a prototype that demonstrates real-time detection and remediation of injection attacks without human intervention.
Poster:	View Poster #19 (PDF)
Judges Scoring:	Submit Evaluation (judges only)

Project # 19Adversarial Repair: Enhancing AI Security via Executable Multi-Agent Debate