Abstract: Techniques for controlling distribution of information from a secure domain by automatically detecting outgoing messages which violate security policies corresponding to the secure domain. Semantic models are constructed for one or more message categories and for the outgoing messages. The semantic model of an outgoing message is compared with the semantic models of the message categories to determine a degree of similarity between the semantic models. The outgoing message is classified based on the degree of similarity obtained from the comparison. A determination is made, based on the classification of the outgoing message, if distribution of the outgoing message would violate a security policy for the secure domain. Distribution of the outgoing message is allowed if no security policy is violated.