Threat Actors Use AI-Generated Summaries to Deliver Ransomware Payloads

This report describes a novel AI-focused social-engineering technique—Invisible Prompt Injection and Prompt Overdose—where attackers hide repeated malicious directives in HTML (zero-width characters, white-on-white text, tiny fonts, off-screen CSS) so email clients, browser extensions, and productivity summarizers output attacker-controlled ransomware deployment steps (including Base64-encoded PowerShell). It explains the attack chain (payload embedding, directive steering, context domination), demonstrates proof-of-concept outputs, assesses impact and risk (scalability, bypassing visual inspection), and recommends mitigations such as client-side sanitization, prompt filtering, payload pattern recognition, context-window balancing, UX safeguards, and enterprise policy enforcement.