subsets of groups which have to be closed no matter what One example of a subset of a group $G$ which has to be closed in any topology on $G$ compatible with the group operations is a centraliser. Are there any other interesting examples?
 A: This is not really an answer, rather a longer (and not very deep) remark about the question of the topology and the assumption that it should be Hausdorff.
Not every topology that makes a group a topological group is Hausdorff. But to any topological group we can associate a Hausdorff topological group in a canonical way. Let $G$ be a topological group and denote by $H$ the closure of $\{e\}$. Then $H$ is a normal subgroup in $G$, and the quotient group $G/H$ is Hausdorff with respect to the quotient topology. See Proposition 1-4 (vi) on page 6 of Ramakrishnan and Valenza, Fourier analysis on number field, 1999.
Based on this result, Ramakrishnan and Valenza write "Part (vi) shows that every topological group projects by a continuous homomorphism onto a topological group with Hausdorff topology. In this sense the assumption that a given group is Hausdorff is not too serious."
Nonetheless, the assumption plays an important role in Rupert's question. For, if we take the trivial topology $\mathcal{O}=\{\emptyset,G\}$, the $H=G$ and $G/H$ is the trivial group. Btw, the example of the trivial topology shows that the answer to the question (if we don't require Hausdorff) is easy: $G$ and the empty set are the only subsets that are closed in any topology on $G$ that makes $G$ a topological group.
A: Subsets of a group that are closed with respect to any Hausdorff group topology
are called unconditionally closed. 
Clearly, all algebraic sets are unconditionally closed, where a subset of a group $G$ is called algebraic if it is an intersection of finite unions of the sets of solutions to some equations with coefficient from $G$.  
A.A.Markov proved that for countable groups the converse is also true:
$$
\hbox{unconditionally closed = algebraic}.
$$
For uncountable groups, this is not always the case as follows from works of S. Shelah (under CH) and G. Hesse. 
