AI pentesting uses reasoning-capable AI models to autonomously find, exploit, and validate context-dependent vulnerabilities — like broken authorization and business-logic flaws — that traditional scanners miss. Unlike DAST tools that match known patterns, AI pentesters reason about intended application behavior and try to subvert it continuously. The approach relies on four components: a reasoning model, deterministic tooling, an independent validator, and target context. Key limitations include inconsistency on hard tasks (61% benchmark success overall, dropping below 26% on harder challenges) and ~30% hallucination rates in raw AI tooling, making independent validation essential. AI pentesting complements rather than replaces DAST and human testers, covering the ~350 days per year that annual manual pentests don't. Six evaluation criteria are provided: independent validation, context awareness, continuous testing, verifiable output, low false positives, and complementarity with existing tools.
