An Adversarial Testing Framework for Multi-Agent Red–Blue Systems in Automated Software Hardening

Software testing has advanced significantly, yet most automated methods remain limited to rule-based coverage rather than adversarial resilience. This paper introduces a Multi-Agent Adversarial Testing Framework (ATF) that simulates a Red–Blue team dynamic to continuously discover and patch vulnerabilities in software systems. Red Team agents generate targeted attacks and edge cases using adversarial reasoning and fuzzing. Blue Team agents respond with automated patch generation, refactoring, and test reinforcement. A Judge Agent evaluates both attack effectiveness and defense quality, forming a continuous self-improvement cycle. Our implementation integrates a large lan- guage model (LLM)-based reasoning with reinforcement learning and static analysis pipelines. Preliminary experiments on open- source projects show a 47% increase in unique bug discovery and a 33% reduction in recurring vulnerabilities compared with baseline fuzzers. These results suggest that adversarial multi-agent systems can significantly advance the robustness and adaptability of automated software testing.