Huge kriging weights when using kriging with external drift in gstat

Dear  all,
I am interpolating automatically  daily air temperature values using 
kriging with external drift (altitude) and the library gstat. From time 
to time (quite random) I get unrealistic predictions like -400?C with 
very high kriging weights   like 2000.
  Looking at the data everything seems to be ok.  Then I changed the 
variogram ranges  a little bit and the problem disappears.
How can this behavior be explained?

  I  have attached a small  example script for demonstration  and some data.

Thanks you
Tobias
-------------- next part --------------
A non-text attachment was scrubbed...
Name: example.R
Type: application/x-extension-r
Size: 1236 bytes
Desc: not available
URL: <https://stat.ethz.ch/pipermail/r-sig-geo/attachments/20130702/1290aff5/attachment.bin>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: points2.Data
Type: application/octet-stream
Size: 865 bytes
Desc: not available
URL: <https://stat.ethz.ch/pipermail/r-sig-geo/attachments/20130702/1290aff5/attachment.obj>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: mypoint.Data
Type: application/octet-stream
Size: 343 bytes
Desc: not available
URL: <https://stat.ethz.ch/pipermail/r-sig-geo/attachments/20130702/1290aff5/attachment-0001.obj>