Definition
Cross-Domain Prompt Injection Attacks (XPIA) occur when attacker-controlled content from one trust domain (e.g., an external Word document) influences an LLM agent’s behavior in another domain (e.g., an enterprise Copilot drafting session).
Key Points
- Extended by hakon-maloy from single-interaction compromise to self-propagating document carriers (ai-worm)
- Architectural: untrusted data shares context window with instructions
- Mitigations that target specific payloads often fail the broader class
Related
- prompt-injection
- ai-worm
- microsoft-copilot
- document-borne-prompt-injection
- enterprise-llm-document-integrity