AI Safety in a World of Vulnerable Machine Learning Systems — LessWrong