Currently, machine learning (ML) systems have impressive performance but can behave in unexpected ways. These systems latch onto unintuitive patterns and are easily compromised, a source of grave concern for deployed ML in settings such as healthcare, security, and autonomous driving. In this talk, I will discuss how we can redesign the core ML pipeline to create reliable systems. First, I will show how to train provably robust models, which enables formal robustness guarantees for complex deep networks. Next, I will demonstrate how to make ML models more debuggable. This amplifies our ability to diagnose failure modes, such as hidden biases or spurious correlations. To conclude, I will discuss how we can build upon this “reliability stack” to enable broader robustness requirements, and develop new primitives that make ML debuggable by design.