Preprint. This study analyses extreme co-movement among thirteen large cryptocurrencies using rolling windows spanning late 2021 to 2025. Dynamic graphical models distinguish direct links during joint crashes and rallies from connections mediated by other assets. The estimated crash network remains dense and stable, whereas the rally network becomes sparser and more sectoral. Comparisons with a Gaussian model show that ordinary co-movement does not adequately describe the dependence observed during extreme market conditions.