Security Degradation in Iterative AI Code Generation -- A Systematic Analysis of the Paradox