The paper discusses selected issues related to the implementation and deployment of the Web Application Firewall that protects the target application by verifying the incoming requests and their parameters through matching them against recorded usage patterns. These patterns in turn are learned from the traffic generated by the users of the application. Since many web applications, including these operated by the government, are prone to exploits, there is a need to introduce new easily implementable methods of protection to prevent unauthorized access to sensitive data. A Learning Web Application Firewall offers a flexible, application-tailored, yet easy to deploy solution. There are certain concerns, however, regarding the classification of data that is used for the learning process which can, in certain cases, impair the firewall ability to classify traffic correctly. These concerns are discussed on the basis of reference implementation prepared by the authors.