New physics at the LHC would typically manifest as an anomalous overdensity of events in some phase space region of the high-dimensional feature space of LHC data. The traditional way to search for new physics is to make some theory-motivated guess as to what it will look like, and then make a phase space selection which is optimized using simulated data and then look in that region for an excess in the real LHC data. Higher sensitivity is often achieved at the expense of introducing stronger assumptions about the underlying signal model, which are used to make more optimised multivariate cuts using more event features. I will discuss a case study of an alternate paradigm, in which sensitive multivariate selections can be be found while maintaining few signal-model assumptions and without the need for potentially unreliable signal simulations. The key ingredient is a machine learning algorithm which searches for event over-densities on an otherwise smooth background, as is often the case in bump hunts for particle resonances. In this 'CWoLa-hunting' (Classification Without Labels) strategy, the selection cuts are not determined in advance but are rather dictated by the distribution of the actual measured LHC data. I will also provide a summary of some of the other ideas for using machine learning for model-agnostic searches that have been proposed in 2018.