Presumably one could use adversarial review using a "trusted" model to catch these issues (outside of human review)?

Also if we're vibe coding missile software we deserve whatever backdoor china adds