Indirect injection plants instructions in content the agent reads on its own, such as web pages, emails, documents, issues, calendar invites, or tool results. The user often never sees the payload, which makes this the primary threat for any agent that touches untrusted data.
Read in this order
- OfficialHow Microsoft defends against indirect prompt injection attacks (MSRC)Start here: a defender's view of the threat and the layered controls used against it.
- PaperNot what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection (Greshake et al.)The original paper that showed these attacks working against real integrated apps.
- PaperInvitation Is All You Need! Promptware Attacks Against LLM-Powered Assistants in ProductionA recent case of calendar invites hijacking production assistants, with the full attack chains.