From the latest Nature (Vol 443, No 7114, p 1013):
When a new, independent code is used for the calculations on which the conclusions of this Letter were based, the results reported for the evolution of obliquity cannot be reproduced. This code was written in the inertial frame and is more reliable than the one used in the Letter. In most runs, the obliquities can change by only a few degrees and attain large values in only a very few cases. In addition, the obliquity variation shown in the Supplementary Information, although correct, originates from changes to orbital inclination of the planet, and close encounters are not effective in causing large obliquities.
In other words, the code was flaky, so the results we published were wrong. Kudos to the author for trying to verify his result with another program—I'm sure a lot of computational scientists would be very embarrassed if they had to do the same. But I wonder: why does he believe the "new code" is more reliable? Surely not just because it uses a more sophisticated method, or gives a less surprising answer...
(Thanks to Andrew Straw for the pointer.)