The simulator will only predict choice for product scenarios you tell it to do. Since having all "feature not included" levels doesn't represent a viable product concept, you shouldn't actually specify it as a product within a simulation scenario.
This issue of using conjoint analysis to deal with a series of binary (on/off) attributes was the topic of Kevin Lattery's presentation at the recent Sawtooth Software conference (October in California). He pointed out that regular conjoint analysis (with main effects estimation) will usually over-predict the choice of products with most of the items "on" and will usually under-predict the choice of products with most of the items "off".
Interaction effects (between attributes taken 2 at a time) can help reduce the modeling problems, but even that cannot do enough sometimes. Kevin proposed a pretty fancy nested logit approach to solving the problem.
But, mere mortals may need a more practical approach that is possible within Sawtooth Software. One fairly doable (with some data re-arranging in the .csv or .cho file, after data collection) approach for your problem is to collapse all 6 attributes into a single attribute with 64 levels, representing all possible combinations of the 6 binary attributes, and recode your data file to reflect that. To help stabilize the estimation under CBC/HB, you can impose a series of utility constraints. For example, having feature A+B should be higher in utility than having feature A or feature B alone, etc. There are dozens of such utility constraints you can impose to ensure better convergence and better precision at the individual-level.
A mixed approach to this is, for example, to consider grouping the 4 binary items that tend to have the most synergies or substitution effects among them and to collapse those 4 binary attributes into a single 16-level attribute. The other two attributes you could collapse into a 4-level attribute. Then, impose utility constraints within CBC/HB estimation, as described in the previous paragraph.