We describe the detection, interpretation, and removal of the signal resulting from interactions of high energy particles with the Planck High Frequency Instrument (HFI). There are two types of interactions: heating of the 0.1 K bolometer plate; and glitches in each detector time stream. The transient responses to detector glitch shapes are not simple single- pole exponential decays and fall into three families. The glitch shape for each family has been characterized empirically in flight data and these shapes have been used to remove glitches from the detector time streams. The spectrum of the count rate per unit energy is computed for each family and a correspondence is made to the location on the detector of the particle hit. Most of the detected glitches are from Galactic protons incident on the die frame supporting the micro- machined bolometric detectors. In the Planck orbit at L2, the particle flux is around 5 cm 2 s 1 and is dominated by protons incident on the spacecraft with energy > 39 MeV, at a rate of typically one event per second per detector. Di ff erent categories of glitches have di ff erent signatures in the time stream. Two of the glitch types have a low amplitude component that decays over nearly 1 s. This component produces excess noise if not properly removed from the time- ordered data. We have used a glitch detection and subtraction method based on the joint fit of population templates. The application of this novel glitch subtraction method removes excess noise from the time streams. Using realistic simulations, we find that this method does not introduce signal bias into the Planck data.