This study focuses on automatically detecting wrong implementations of specification in Java Card programs, without any knowledge on the source code or the specification itself. To achieve this, an approach based on Natural Language Processing and machine-learning is proposed. First, an oracle gathering methods with similar semantics in groups, is created. This focuses on evaluating our approach performances during the neighborhood discovery. Based on the groups automatically retrieved, the anomaly detection is based on Control Flow Graph of programs of these groups. In order to benchmark its ability to detect vulnerabilities, another oracle of vulnerabilities is created. This oracle knows every anomaly the approach should automatically retrieve. Both the neighborhood discovery and the anomaly detection are benchmarked using the precision, the recall and the F1 score metrics. Our approach is implemented in a tool: Confiance and it is compared to another machine-learning tool for automatic vulnerability detection. The results expose the better performances of Confiance over another approach in order to detect vulnerabilities in open-source programs available online.