Multi-Agent Harness Synthesis Automates Vulnerability Discovery That Human Auditors Missed for Decades
arXiv·high signal
Researchers present a framework that automatically synthesizes multi-agent harnesses — programs that wire together specialized LLM agents for security auditing — to find real vulnerabilities in source-available targets. The system splits work among discovery, exploitation, and verification agents, finding bugs that human auditors and automated fuzzers missed for decades. This directly addresses the growing need for autonomous security testing as codebases outpace manual review capacity.