indirect prompt injection
Prompt injection hidden in content the model reads on someone else's behalf, such as a web page, email, ticket, document or tool result.
The attacker reaches the user's data and tools without ever interacting with the application.
In more detail
Greshake and colleagues showed in 2023 that an attacker never needs to talk to the model: they plant instructions where it will read them. It is the dangerous kind for agents, because the victim is the user whose data and tools the model holds.
Related terms
- prompt injection — An attack in which text a model reads overrides or redirects the instructions it was given, because instructions and data share one channel.
- lethal trifecta — Simon Willison's name for an agent that has access to private data, exposure to untrusted content and a way to communicate externally, which together let injected text steal data.
- spotlighting — Marking untrusted text in a prompt, by delimiting, interleaving a marker or encoding it, so the model can tell it is data and not instructions.
Taught in
- Prompt Injection — AI Engineering: LLM Apps, RAG & Agents
References
- Greshake et al. (2023): Not what you've signed up for: indirect prompt injection
- OWASP GenAI Security Project: LLM01 Prompt Injection