Source: The Dual LLM pattern for building AI assistants that can resist prompt injection

Simon Willison (2023): split an assistant into a tool-wielding privileged LLM that never sees untrusted text and a tool-less quarantined LLM, coordinated by plain code.

Contents

Summary

Proposes an architecture for an assistant that can act on private data without obeying injected instructions. A privileged LLM handles trusted user input and has tools; a quarantined LLM processes untrusted content and has no tools; a controller (ordinary code) runs both and holds untrusted text in variables. The privileged LLM refers to content only by symbolic names such as $VAR1, so attacker text never enters its context.

Key claims

  • Quarantined output must never flow back to the privileged LLM, except when constrained to a validated fixed set of values.
  • The author calls the design complex and harmful to UX, and warns that later features can silently reintroduce leaks.
  • It does not stop social engineering of the user (e.g. getting them to paste obfuscated data out).
  • A 2025 update notes that Google DeepMind’s CaMeL paper found a flaw (data flow can still be manipulated) and extends the idea.

Feeds

Prompt injection

Suggest changes →