Interruptr

Multi-Agent C/C++ Vulnerability Detection

Interruptr is a multi-agent system for automated C/C++ vulnerability detection that orchestrates specialized LLM agents to analyze source code at different levels of abstraction. The system achieves approximately 96% vulnerability detection rate on standard benchmarks, dramatically outperforming Cppcheck's 12% detection rate while reducing cost by 95% compared to direct GPT-4 analysis.

The agent architecture consists of specialized roles: a Lexical Agent for pattern matching and known vulnerability signatures, a Semantic Agent for data flow and taint analysis, a Contextual Agent for understanding program logic and intent, and a Synthesizer Agent that combines findings and eliminates false positives. This division of labor allows using smaller, cheaper models for routine analysis while reserving expensive models for complex cases.

Built on AutoGen and LangGraph frameworks, the system supports configurable agent topologies -- from simple sequential pipelines for fast scanning to full adversarial debate for high-assurance analysis. The conversation protocol includes structured vulnerability reports with CWE classification, severity scoring, and suggested remediation code.

This work has been submitted to COMPSAC 2026 (IEEE International Conference on Computers, Software, and Applications), demonstrating the cost-quality Pareto frontier of multi-agent vulnerability analysis.

• ~96% vulnerability detection rate vs Cppcheck baseline of 12%
• 95% cost reduction compared to direct GPT-4 analysis
• Submitted to COMPSAC 2026
• Multi-agent architecture with Lexical, Semantic, Contextual, and Synthesizer agents
• Built on AutoGen and LangGraph with configurable analysis topologies