What happens in estimation for unacceptables in ACBC is that it typically (if you follow default procedures) only asks a respondent if a level is unacceptable if it has been consistently rejected in the Screening phase of the questionnaire to that point. Then, if a respondent marks a level as unacceptable, then the software looks forward to the concepts not yet seen by the respondent and marks any concepts containing an unacceptable level as "not a possibility".
So, the utilities are just estimated by fusing (stacking) the BYO, Screener, and Tournament section choice tasks. If a respondent marks a level as unacceptable, then every time a concept containing this level was shown (or was going to be shown) in the Screener section for this respondent, it will have been marked "not a possibility."
Also, remember that HB involves some Bayesian smoothing toward population means. So, if a respondent is the only respondent to mark a level as unacceptable, then it's quite possible that the bulk of the population preferences will tend to shift that utility weight more positive than you might have expected by just looking at that respondent's choices.
You should not try to compare the utility of levels from different attributes. So, don't try to compare the unacceptable utility for a level of one attribute to the unacceptable utility for a level of a different attribute. You can only compare utilities within attributes.