We used the multiphase and multicomponent TOUGH2/EOS7CA model to carry out predictive simulations of CO2 injection into the shallow subsurface of an agricultural field in Bozeman, Montana. The purpose of the simulations was to inform the choice of CO2 injection rate and design of monitoring and detection activities for a CO2 release experiment. The release experiment configuration consists of a long horizontal well (70 m) installed at a depth of approximately 2.5 m into which CO2 is injected to mimic leakage from a geologic carbon sequestration site through a linear feature such as a fault. We estimated the permeability of the soil and cobble layers present at the site by manual inversion of measurements of soil CO2 flux from a vertical-well CO2 release. Based on these estimated permeability values, predictive simulations for the horizontal well showed that CO2 injection just below the water table creates an effective gas-flow pathway through the saturated zone up to the unsaturated zone. Once in the unsaturated zone, CO2 spreads out laterally within the cobble layer, where liquid saturation is relatively low. CO2 also migrates upward into the soil layer through the capillary barrier and seeps out at the ground surface. The simulations predicted a breakthrough time of approximately two days for the 100kg d(-1) injection rate, which also produced a flux within the range desired for testing detection and monitoring approaches. The seepage area produced by the model was approximately five meters wide above the horizontal well, compatible with the detection and monitoring methods tested. For a given flow rate, gas-phase diffusion of CO2 tends to dominate over advection near the ground surface, where the CO2 concentration gradient is large, while advection dominates deeper in the system.