A couple of things:
1. Direct the AI to look specifically into areas where I have a gut feel things might go wrong
2. An assessment by another model (eg Codex reviews Claude's work)
3. Manual testing
I think the key is treating AI-generated code like code from a junior developer: review it carefully, test it thoroughly, and never assume it's correct just because it was generated.
If you are unsure about the AI generated code, it's better to not ship a risky product. You can also use 3rd party services for the most vulnerable areas of your product.
AI says, "A proof is a program that computes the witness."