Structure-based drug design (SBDD) relies on holo structures because ligand binding can expose pockets absent from apo structures. Yet holo structures are difficult to obtain, requiring a suitable ligand and experimental capture of the bound state. This limitation is especially consequential for proteins such as KRAS, whose apo structure lacks obvious binding pockets and was long considered “undruggable.” Nevertheless, apparently featureless proteins can expose cryptic pockets—potential drug-binding sites hidden in the apo state—through protein motions or binding-associated changes. Discovery of the hidden switch-II pocket in KRAS G12C, followed by approvals of sotorasib and adagrasib for KRAS G12C-mutated non-small-cell lung cancer, demonstrated that such pockets can become clinically actionable. We present CrypticFlow, an AI framework that generates holo-like conformations from apo structures. It combines flow matching and an E(3)-equivariant graph neural network for coarse generation with hierarchical refinement in local internal-coordinate systems that preserves adjacent Cα geometry. On the D3PM benchmark, CrypticFlow achieved the best apo-to-holo prediction performance among the evaluated models and the strongest Cα-level and cg2all-reconstructed all-atom structural validity. By expanding protein conformations when holo structures are unavailable, CrypticFlow may support virtual screening and lead discovery for challenging targets.