Semantic leakage (or inference leakage) occurs when an AI system reveals sensitive information by inferring or combining it, surfacing a restricted conclusion even though no individually-restricted file was accessed.
What is semantic (inference) leakage?
Because models reason over many inputs, they can assemble a sensitive insight from fragments that were each, on their own, permitted. Access logs show nothing unusual, which is what makes semantic leakage so dangerous and so invisible to DLP and file-level controls. Preventing it requires governing the meaning being assembled and inferred, before the answer is formed.