This page lists known problems and bugs in LAMMPS that have not yet been fixed. Typically this is because the fix is non-trivial and/or the bug is obscure enough it shouldn't cause most people problems. Once the bug is fixed it will disappear from this page and appear as a patch on the New features/Bug fixes page.
Since this is only the list of known bugs, the following quote seems apropos:
As we know, there are known knowns. There are things we know we know. We also know there are known unknowns. That is to say, we know there are some things we do not know. But there are also unknown unknowns, the ones we don't know. -- Donald Rumsfeld
This page lists unknown bugs in LAMMPS.
June 2008
Using non-Coulombic tabulated potentials via pair_style table under pair_style hybrid with other Coulombic potentials and a long-range Kspace solver (PPPM or Ewald) can cause LAMMPS to generate an error if the cutoffs of all the potentials are not the same. This is b/c LAMMPS requires all Coulombic potentials to have the same cutoff and isn't recognizing that the tabulated potentials aren't Coulombic.
Apr 2008
Fix langevin and pair lubricate hardwire the timestep in their calculations. This means they will use an invalid timestep when used with rRESPA or possibly with the adaptive timestepper fix nve/limit.
Feb 2008
When a system has bonds (with special_bond settings != 1), and the pair cutoff > box size, the image contributions to the pair interactions are not correct. Things are fine if no bonds are present or special_bonds = 1 1 1.