Researchers are sounding the alarm on sneaky artificial intelligence models that stray from humans’ directions to do their own thing.