So could someone, say, cajole the AI to output a fuzzer with the goal of finding a path to executing arbitrary code?

If so, then any similar user-input powered AI applications might run this kind of risk.