As mentioned in the comments, your NFA actually accepts all strings. What you have hit upon is the fact that NFAs are not resilient to complementation. Whereas given a DFA for a language, you can turn it to a DFA for the complement of the language by complementing the set of accepting states, the same doesn't hold for NFAs. In fact, "complementing" an NFA (that is, constructing an NFA for the complemented language) could result in exponential blow-up in the number of states!
As an example, consider the language of all words over $\{1,\ldots,n\}$ which do not contain all symbols. There is an NFA with $n$ states accepting the language (the NFA has multiple initial states, which might not be allowed under your definition), but any NFA for the complement must contain at least $2^n$ states. This blow-up is worst possible due to the powerset construction which converts an NFA with $n$ states to an equivalent DFA with $2^n$ states, which can be complemented without increasing the number of states.