Typically, the reason that WM and CSF are used as nuisance regressors in connectivity analyses is because these types of analyses are looking at correlations in the time-series obtained from different parts of the brain, and correlations of WM/CSF signals between two brain region can introduce bias into these measurements (e.g. producing artifactual correlations). In a standard GLM analysis of task data, the focus is on the the change in HRF magnitude based on stimuli/task factors. In this case, correlations across voxels related to WM/CSF signals are unlikely to produce a bias in the HRF magnitude estimates, and so they are typically not regressed out of the data in task-based GLM analyses. Depending on your smoothing kernel and precision of template normalization, you might want to use a grey matter mask as an inclusive mask to avoid conducting tests at the thousands of WM/CSF voxels (and thereby incurring a stricter multiple comparison correction), but this is not necessary per se. 

In task-based designs you are using signal averaging to get around much of the concerns in resting state data. Motion is much less of an issue.Still software such as FSL allows you to include head motion regressors as covariates as well downweight frames based upon motion. There is a very nice paper by Josh Segel looking at frame censoring in task-fMRI though. This may be something you could take a look at as well.

http://onlinelibrary.wiley.com/doi/10.1002/hbm.22307/full

As previous members have pointed out, these nuisance regressors present more of a problem with correlating low-freq oscillations in rs-fMRI and not much of a factor when correlating with a task design. You are more likely to see an improvement in task-based studies from methods to remove multiple sources of subject-specific noise, scanner noise, and thermal noise, such as with ICA denoising.

On a related note, while frame censoring is advised in resting state data, you must exercise caution in task-based studies. When the motion is correlated with the task, you can have significant negative effects on your data. (See http://www.personal.reading.ac.uk/~sxs07itj/Johnstone_HBM_Oct_2006.pdf).

Again, my experience is that you generally have more significant improvement with techniques such as ICA denoising.

http://www.personal.reading.ac.uk/~sxs07itj/Johnstone_HBM_Oct_2006.pdf

Thanks for the answers. ICA-denoising (e.g. via ICA-AROMA) is always a good choice or at least a good try :) ... I also thought that WM and CSF nuisance regressors are particularly important for rs-fMRI, but less so for task-based studies, but wasn't sure about the latter since there are some task-based studies that included them. But I agree that there are more important techniques to get rid of noise or explain additional variance in task-based studies. 

PS: I also think that frame censoring/scrubbing can be problematic (e.g., reduced network reproducibility, reduced temporal degrees of freedom) and that ICA-denoising seems to be the better choice, see http://www.ncbi.nlm.nih.gov/pubmed/25770990