Undesirable natural aging (NA) in Al-6xxx delays subsequent artificial aging (AA) but the size, composition, and evolution of clustering are challenging to measure. Here, atomistic details of early-stage clustering in Al-1% Mg-0.6%Si during NA are studied computationally using a chemically-accurate neural-network potential. Feasible growth paths for the preferred beta '' precipitates identify: dominant clusters differing from beta '' motifs; spontaneous vacancy-interstitial formation creating 14-18 solute atom beta ''-like motifs; and lower-energy clusters requiring chemical re-arrangement to form beta '' nuclei. Quasion-lattice kinetic Monte Carlo simulations reveal that 8-14 solute atom clusters form within 1000 s but that growth slows considerably due to vacancy trapping inside clusters, with trapping energies of 0.3-0.5 eV. These findings rationalize why cluster growth and alloy hardness saturate during NA, confirm the concept of "vacancy prisons", and suggest why clusters must be dissolved during AA before formation of beta ''. This atomistic understanding of NA may enable design of strategies to mitigate negative effects of NA.