My background is in mechanical engineering. When I transitioned into cybersecurity, I brought a physical mindset with me: a system only works when the right forces are applied to the right mechanisms.
Today, the tech industry tends to treat AI like a magic wand. I see it differently. Large Language Models are a brilliant new technological commodity—much like electricity. You don't ask electricity to build a car; you use it to power the machines that do. You treat it with respect, implement security boundaries, and remember the golden rule of engineering: use the right tool for the right job.
When you apply this to automated security testing and agentic workflows, the current approach is fundamentally broken. We are asking language models to do precision machining. Agents get stuck in trial-and-error loops—guessing payload syntaxes, hitting CLI flag errors, and hallucinating schema parameters. It burns massive amounts of execution credits. Even worse, when an agent finally produces a functional template, it often causes state-mutation friction. It modifies application settings or pollutes enterprise databases without cleaning up, forcing maintainers to reject the pull requests to protect production environments.
