Prompt injection attacks don't look like what you’re seeing in social media and headlines · Blog · Sublime Security

This report explains how attackers use indirect prompt-injection—hidden or benign-looking text embedded in phishing emails—to influence AI/ML email security systems into misclassifying malicious messages as legitimate. It provides concrete examples (hidden newsletter copy and near-background-colored text from a novel), notes detection by Sublime's ASA, and argues that such nuanced evasion techniques will continue to evolve, requiring continual adaptation of defensive AI.