An AI worm is a worm-like propagation pattern in which model-generated content carries malicious instructions forward into later files, messages, or workflows. Unlike traditional malware, it does not need to execute code on its own. Instead, it survives by being copied, rewritten, or reused by an AI assistant that treats hidden instructions as legitimate guidance.
This matters because AI features in office tools, chat systems, and file generators can blur the line between data and control. Attackers may hide prompt-injection payloads in documents or other content so the model repeats them into a new draft, template, or shared artifact. Defenses focus on strong trust boundaries: treat document text as untrusted input, inspect or strip hidden content and metadata, limit what the model can carry forward, and require human review before generated output is reused downstream.