Researchers at UC Santa Barbara tested SAILOR, a structured pipeline combining CodeQL static analysis, LLM-driven symbolic execution harness generation, and runtime validation, on ten open-source C/C++ projects and found 379 previously unknown memory-safety vulnerabilities — roughly 30 times more than Claude Code achieved with unlimited autonomous access to the same codebases. The finding suggests that constraining AI tools to specific, sequential tasks with human-defined boundaries (static analysis → harness synthesis → proof) outperforms unconstrained agent-style vulnerability hunting. This matters because it demonstrates a scalable approach to autonomous vulnerability discovery that could accelerate patch cycles but also raises operational security concerns about the speed of automated exploitation pipeline development.