Perhaps one way to mitigate this is by wrapping the calls to another system such that all trading is supervised.

Another technique is to provide API keys in the wrapper so that test programs will not have the keys to a live system.

The real problem is that the risk of these test systems have not been sufficiently identified or recognized. We are all too busy creating mock systems instead of devoting sufficient oversight to the development of test software.