Web App Pentesting in the AI Era
A practitioner’s account of using frontier and open-weight LLMs for web application penetration testing, contrasting lazy inference (single “find all vulns” prompts) against structured harnesses that enforce thoroughness and execution-grounded validation. The author decomposed AI systems into model, inference engine, and harness layers, observing that models like Claude Opus 4.7 are technically capable of finding real vulnerabilities but stop prematurely without iterative prompting or loops over individual code files. Key technical insight: models confidently assert findings without proof; validation requires forcing repeatable proof-of-concept payloads against running targets rather than accepting model-reasoned claims. The article demonstrates how specificity and execution-grounded evidence significantly improve both thoroughness and accuracy in AI-assisted security testing, offering practical methodology for security professionals integrating AI into their workflow.