RL finds compact QKD attacks that nearly hit dynamic-programming bounds under noise and drift, because static eavesdropping was too easy.