In a search for short-time-scale astrophysical transients in time-domain data, radio-frequency interference (RFI) causes both large quantities of false positive candidates and a significant reduction in sensitivity if not correctly mitigated. Here, we propose an algorithm that infers a time-variable frequency channel mask directly from short-duration (∼1 s) data blocks: the method consists of computing a spectral statistic that correlates well with the presence of RFI, and then finding high outliers among the resulting values. For the latter task, we propose an outlier detection algorithm called Inter-Quartile Range Mitigation (IQRM), which is both non-parametric and robust to the presence of a trend in sequential data. The method requires no training and can, in principle, adapt to any telescope and RFI environment; its efficiency is shown on data from both the MeerKAT and Lovell 76-m radio telescopes. IQRM is fast enough to be used in a streaming search and has been integrated into the MeerTRAP real-time transient search pipeline. Open-source PYTHON and C++ implementations are also provided.