We report the detection of a dark substructure through direct gravitational imaging - undetected in the HST-ACS F814W image - in the gravitational lens galaxy of SLACS SDSSJ0946+1006 (the Double Einstein Ring). The detection is based on a Bayesian grid reconstruction of the two-dimensional surface density of the galaxy inside an annulus around its Einstein radius (few kpc). [...] We confirm this detection by modeling the system including a parametric mass model with a tidally truncated pseudo-Jaffe density profile; in that case the substructure mass is M_sub=(3.51+-0.15)x10^9 Msun, located at (-0.651+-0.038,1.040+-0.034), precisely where also the surface density map shows a strong convergence peak. [...] We set a lower limit of (M/L)_V}>=120 (Msun/L}_V,sun (3-sigma) inside a sphere of 0.3 kpc centred on the substructure (r_tidal=1.1kpc). The result is robust under substantial changes in the model and the data-set (e.g. PSF, pixel number and scale, source and potential regularization, rotations and galaxy subtraction). Despite being at the limits of detectability, it can therefore not be attributed to obvious systematic effects. Our detection implies a dark matter mass fraction at the radius of the inner Einstein ring of f_CDM=2.15^{+2.05}_{-1.25} percent (68 percent C.L) in the mass range 4x10^6 Msun to 4x10^9 Msun assuming alpha=1.9+-0.1 (with dN/dm ~ m^-alpha). Assuming a flat prior on alpha, between 1.0 and 3.0, increases this to f_CDM=2.56^{+3.26}_{-1.50} percent (68 percent C.L). The likelihood ratio is 0.51 between our best value (f_CDM=0.0215) and that from simulations (f_sim=0.003). Hence the inferred mass fraction, admittedly based on a single lens system, is large but still consistent with predictions. [...]