Glossary
AI worm
A prompt injection that tells the AI reading it to copy it into what the AI writes next (an email, a file, a chat message), so it spreads from agent to agent.
Researchers demonstrated one in 2024 (Morris II). In 2026 OpenAI reported that its own red-teaming models wrote injections that spread through email, files and Slack in training environments. Checking what agents send, not only what they read, slows the spread.