WeaknessASC-2026-0027Other weakness
Self-replicating prompt injection (AI worms)
MediumWorkarounds only
Injected text that tells an agent to copy it into its own outputs (emails, files, comments) spreads to every agent that reads those outputs next.
How to detect it
- Outgoing messages or files that contain instructions copied from incoming content
- The same unusual text appearing across many agent outputs
Workarounds
- Scan agent outputs as well as inputs for instruction-like text
- Require approval before agents send messages or write shared files
- Rate limit agent-generated outbound messages
Demonstrated in research (the Morris II worm, 2024), in 2026 against Copilot in Word, where hidden text copied itself into each new document, and by OpenAI's red-teaming models, which wrote injections that spread through email, files and Slack in internal research checkpoints (disclosed September 2026); defenses are the same as for prompt injection.
How to investigate
- Trace the first message that carried the text and every agent output that copied it
- Quarantine affected mailboxes, channels or files
Sources
- Cohen, Bitton and Nassi: Here Comes The AI Worm: Unleashing Zero-click Worms that Target GenAI-Powered Applications (2024)
- Håkon Måløy: Context Collapse, Part 3: AI Worming through Word (July 2026)
- OpenAI: Self-replicating prompt injections exist (September 2026)
- OWASP Top 10 for LLM Applications 2025: LLM01 Prompt Injection
- OWASP Top 10 for Agentic Applications 2026: ASI08 Cascading Failures
- MITRE ATLAS: AML.T0061 LLM Prompt Self-Replication