All good until we get a new vector of subtly poisoned training data introducing vulnerabilities if one doesn't understand the generated code.
This form of argument appears in every thread about LLMs: They are fallible, what will we ever do?!
As if people haven't been fallible in both the same and different ways all along. How have we ever managed...
Others will come to do the same. Some others will not.
Not sure that your generalization is true here.