The setup: Proofpoint's email protection system scores every incoming email and labels it spam if the score is too low. Those scores were visible in email headers.

The attack: Researchers Will Pearce and Nick Landers collected enough of these scores to train a “copy-cat” machine learning model that mimicked Proofpoint's classifier — without ever touching Proofpoint's source code or training data directly.

The result: With a working copy of the model's decision logic, attackers could craft malicious emails specifically engineered to score well and slip past the real filter.