Overview
Integrity risks when enterprise productivity LLMs (e.g. microsoft-copilot for Word) ingest untrusted or carrier documents that can alter outputs and self-propagate prompt-injection payloads.
Timeline
- 2026-07-28/29: hakon-maloy AI worm disclosure after 144-day coordination (2026-08-01-microsoft-copilot-word-prompt-injection-worm)
Key Players
Analysis
Class remains open after model-upgrade mitigations — treat as architectural xpia problem, not a one-off bug. Hygiene: limit Copilot ingestion of external docs; assume internally created docs can become carriers.