There is also this paper that says that this AI loop makes the security worse getting almost 40% rise of vulns after just 5 iterations https://arxiv.org/html/2506.11022v2
An agent reviewing its own output has the same blind spots, the same training-data gaps, and a strong bias toward declaring its work "good". But even having a different model review the code is flawed and not good enough, especially for those whose code can't be wrong, like in regulated industries.