Adversarial prompt injection payload for evading AI-based detection · Blog · Sublime Security

This Attack Spotlight describes a real-world credential phishing campaign that combines standard phishing lures (fake document-share links, urgency, redirect chains, and hijacked threads) with hidden prompt-injection payloads aimed at AI security scanners. The attackers embed fabricated Edge tab metadata and first-person AI phrasing in the email HTML to poison context, attempt self-negation of analysis, and evade detection, while the visible phishing flow ultimately redirects victims to credential-harvesting pages.