Indirect Prompt Injection
A security vulnerability where malicious instructions are hidden within external data that a language model ingests, hijacking its intended behavior.
Think of It Like This
Like slipping a secret note into a textbook that tricks the reader into performing a prank when they read that specific page.
This occurs when an LLM processes untrusted content, such as summarizing a web page containing hidden commands. The model unwittingly executes these injected instructions, potentially leaking user data or generating harmful outputs. Defending against this requires robust input sanitization and strict separation of instructions from data.