The ideal cipher model is a way of modeling of block cipher (i.e. a keyed permutation family) which is very close to the modelization of a hash function by a random oracle. In fact, these two models are even equivalent (see http://arxiv.org/abs/1011.1264).
Recall that a random oracle is a "magic box" with for any new input outputs a purely random value (for repeated inputs, the oracle just repeats itself). A ideal cipher behave similarly given a new input pair (K,x), the ideal cipher outputs a random value y, with a small caveat: instead of being chosen at random for a fixed set, y is chosen among the values which have not yet been used for the key K.
Ideal ciphers can appear in security proofs to show that a mode of operation is secure with a "perfect" blockcipher. In particular, when working in the ideal cipher model, you implicitely assume that your blockcipher is secure against related-key attacks.
However, as with the random oracle model, it is possible to construct cryptographic primitives (albeit artificial ones), which are secure in the ideal cipher model and insecure when instantiated with any real blockcipher. For this reason, proofs in the regular security model (where the blockcipher keyed with a random key is indistinguishable from a random permutation) are considered to be much more representative of real-life security.
If you are interested by proofs in the ideal models (RO or ideal cipher), look for papers about indifferentiability.