Changepoint Detector can look for a change in a series' level and volatility. On heavy-tailed noise (t with 3 degrees of freedom) the Gaussian version reported a break 37 to 63% of the time when none existed. The --robust Laplace cost cut that to 5 to 8%, at a price: on clean Gaussian data with 100 points it found 57% of real breaks against 94%.
Until now only the command line could switch it on. The page has a checkbox for it, shown with that trade-off in its label, and it only applies when the model is "Level and volatility". Still open: a cost that copes with GARCH-style noise, where Laplace still leaves 21 to 29% false alarms, and the Bayesian view.