Dfinity has a 2/3 honesty assumption (same as Ethereum 2.0 for notarisation) and targets a failure probability \rho=2^{-40}. (See section 4.2 “Threat Model” of their consensus whitepaper. When \beta=3 and \log_2\rho=-40 they get a minimum committee size of 423.)
To reach their failure probability for liveness (which they call “availability”) Dfinity assumes that 100% of their honest nodes always eventually come back online. In effect, Dfinity has a hidden “100% eventual liveness” assumption.
In the context of WW3 (for which we want Ethereum 2.0 to survive) let’s assume that just 15% of the honest nodes are permanently offline (e.g. an atomic bomb wipes out Palo Alto). Now only 56.66% of the network is both honest and live and the failure probability goes above 2^{-9}. Depending on how fast committees are refreshed, I expect it would take just a few days to get a bad committee that can stall the network permanently.
(Side note: I’m a fan of Dfinity and believe their blockchain will work well in the optimistic “no long-lasting catastrophies” setting.)