LLM Prompt Injection Worm

Researchers presented ComPromptMized (Morris II), a proof-of-concept worm that leverages adversarial, self-replicating prompts embedded in text or images to hijack GenAI-powered email assistants (via RAG) to exfiltrate data and propagate automatically; the paper demonstrates replication and propagation across models (Gemini Pro, ChatGPT 4.0, LLaVA) under black- and white-box settings and evaluates factors affecting spread and malicious activity.