I think that a powerful application of AI is the breaking of apps. Vibe breaking rather than vibe coding. AI can help immitate nasty users, find edge cases, hack systems, and break them under load. If you have a non-graphical intermediate layer the setup for attempting such attacks with a typical AI is easy; using a GUI with LLMs is possible though harder, so better think of how to make such extreme attacks by AI easy when you start to vibe code.
I was hoping for a bit more on what happens when the project is ran and interacted with. The two-channel correspondence we see between code and behavior that we normally see as congruent must mismatch when vibe coding, no?