Abstract: In one embodiment, detecting unwanted electronic mail messages based on probabilistic analysis of referenced resources comprises receiving a whitelist and a blocklist each having a plurality of network resource identifiers that have appeared in prior messages; retrieving a particular network resource identifier; generating a list of properties for the particular network resource identifier; training a probabilistic filter using the properties; and repeating the retrieving, generating and training for all the network resource identifiers in the whitelist and blocklist. Thereafter, when an electronic mail message is received and contains a URL or other network resource identifier, a spam score or threat score can be generated for the message by testing properties of the network resource identifier using the trained probabilistic filter.