Detecting and blocking content that can develop undesired behavior by artificial intelligence (AI) entities toward users during a learning process is provided. Input information is received for a set of one or more AI entities. Characteristics of the input information are evaluated based on rules of a selected policy from a set of policies and learned characteristics of information associated with a corpus of information. It is determined whether a result of evaluating the characteristics of the input information exceeds a predefined threshold. In response to determining that the result of evaluating the characteristics of the input information exceeds the predefined threshold, the input information for the set of AI entities is filtered by performing a selective filtering action, using a firewall, based on context of the input information.