The logic is Defense in Depth. Even if the "cage" code is AI-written and imperfect, it still creates a barrier. The probability of AI accidentally writing malicious code is high. The probability of it accidentally writing code that bypasses the imperfect protection it wrote itself is much lower