Short answer: If you use superposition of atomic densities to form the initial density matrix, that matrix is not consistent with the geometry of the system, and so the initial energy doesn't mean a lot. As such it is very common for the energy to increase after the first cycle, as it is only then that you get a "meaningful" density matrix for the first time. Here's an example from CRYSTAL:
ijb@ijb-Latitude-5410:~/work/crystal/work/chab$ grep CYC scf.out | grep ETOT | head -3
CYC 0 ETOT(AU) -5.229343032813E+03 DETOT -5.23E+03 tst 0.00E+00 PX 1.00E+00
CYC 1 ETOT(AU) -5.229076341815E+03 DETOT 2.67E-01 tst 0.00E+00 PX 1.00E+00
CYC 2 ETOT(AU) -5.229783696202E+03 DETOT -7.07E-01 tst 4.57E-03 PX 1.66E-01
Slightly longer answer: The initial density matrix formed by superposition of atomic densities (SAD) is block diagonal, each block corresponding to the density matrix for the isolated atom. But in the system the atoms aren't isolated, even at this stage when forming the full density matrix for the system the overlap matrix couples the AOs of the guess on different sites, and in forming the SAD initial density matrix we have neglected the off-diagonal blocks caused by such coupling. Thus the SAD density matrix is "unphysical", thus the energy derived from it has little meaning, thus it is not uncommon for the energy to increase on the first cycle.