In this paper we describe two different DFA attacks on the AES. The first one uses a fault model that induces a fault on only one bit of an intermediate result, hence allowing us to obtain the key by using 50 faulty ciphertexts for an AES-128. The second attack uses a more realistic fault model: we assume that we may induce a fault on a whole byte. For an AES-128, this second attack provides the key by using less than 250 faulty ciphertexts. If we extend our hypothesis by supposing that the attacker can choose the byte affected by the fault, our bit-fault attack requires 35 faulty ciphertexts to obtain the secret key and our byte-fault attack requires only 31 faulty ciphertexts.