Originally, we proposed confirmation scheme to handle data availability problem like this: the block is confirmed(or finalized) if all transaction senders confirm it (and this is done by CAS construction). But it is not only inefficient itself, but also cannot protect non-sender of that block. In the example above, User B cannot safely exit because already the invalid block#1 was withheld by the operator and finalized(note that it only needs User A)
See also this: Plasma EVM 2.0: state-enforceable construction