This is probably a wraparound bug in the 3D FFT-based blurring function. I'd fix it now, but we are in the middle of making vast numbers of other changes to all AFNI codes to make AFNI work with C++. These have to be finalized before anything else can be changed. Finishing this up is probably at least a week away.
One way to deal with this now would be to use the option
-1filter_expr 4 'exp(-r*r/1.44268)' instead of
-1blur_fwhm 2.
bob cox