Why it matters

Agentic AI creates new attack classes. Understanding emerging threats shapes defense.

Advertisement

The architecture

Injection point: infected email, document, or web page.

Propagation: agent processes content, includes malicious payload in future outputs.

AI worm mechanismInfection pointemail/doc/pageAgent processesincludes payloadPropagatesto new targetsDefenses: strict output policies, prevent agent from including untrusted content in outbound
Worm propagation.
Advertisement

How it works end to end

Payload: instructions to include propagation content when writing new content.

Vector: any agent-mediated content sharing (email, chat, doc collaboration).

Defenses: strict output policies, provenance tracking, agent-generated content clearly marked.