The absence of a clear signal of physics beyond the standard model at the CERN LHC motivates search strategies that do not presuppose a specific signal hypothesis. This proceedings present machine-learning-based anomaly detection to search for new physics in a model-agnostic way. The first such search by CMS looks for dijet resonances whose jets have substructure atypical of jets initiated by light quarks or gluons, using five complementary anomaly detection methods applied to 138 fb$^{-1}$ of proton-proton collision data at $\sqrt{s} = 13$ TeV. The technique has been validated directly on data by recovering the top quark without the use of labels. The program has also moved into real time: two unsupervised algorithms, AXOL1TL and CICADA, now run on field-programmable gate arrays inside the CMS level-1 trigger and selected more than four billion collision events during 2024 data taking. Finally, the extension of resonant anomaly detection to event-level observables is outlined.