One or more lock words in a non-volatile memory with write ability correspond to lockable features of a protected system including the memory. A lockable feature should be locked when the corresponding lock word has a value equal to one of a limited number of predetermined locking combination(s). The locking combination(s) are selected so as to minimize the probability of the lock word waking up from manufacture with a value equal to one of the locking combination(s). In order to detect whether allowable usage of a lockable feature should be at a predetermined locked or unlocked level, multi-sampling of the corresponding lock word value is performed. If there is variation among the sampled values of the lock word, a malicious attack is identified. In one preferred embodiment, the multi-sampling occurs upon power up reset of the protected system and if a malicious attack is identified, the protected system is kept in reset.