Dual LLM pattern
dual-LLM · dual LLM architecture
A prompt-injection defense that splits a privileged model, which holds the tools, from a quarantined model, which reads untrusted content and never calls tools; proposed by Simon Willison in 2023.
In our briefings
The claim each of the newest briefings on Dual LLM pattern would have you leave with.
The guide's answer to tool misuse is to give each step of the plan only the tools it needs. The Planner may need no tools at all, and the Executor gets only the tool or tools the current step needs, so that reaching the email tool during a calculation step is meant to be "architecturally impossible". In the authors' own LangGraph code, the planner model names the single tool for each step and the Executor runs only that tool, so the limit comes from the Planner's own output, and the paper does not say what stops an injected or mistaken Planner from naming a broader tool. The authors' own threat table says Plan-then-Execute does not scope tools by itself, so the scoping comes from the frameworks: coded by hand per node in LangGraph, declared per task in CrewAI, where Task.tools overrides Agent.tools (a rule they cite to a community forum post), and set through agent capabilities or custom logic in AutoGen. On top of that they propose RBAC roles as a ceiling, assigned by the Planner at plan time and checked against a predefined permission map, and say CrewAI does not implement this yet; nothing says what limits the role the Planner picks. The paper does not say how an agent authenticates to a tool or which user's access it acts under; credentials appear only as placeholder API keys in its code examples.
A guide to securing agent plans has the planner choose each step's tools and leaves out agent identityBI-FIND-2025-0361
Fixing an agent's plan before it reads untrusted data keeps text in that data from adding actions, and the actions the plan allows can still be steered. In the authors' own example, an assistant asked to send today's schedule to the user's boss commits to reading the calendar and writing one email. Text planted in the calendar "cannot inject any new instructions", they write, but it could arbitrarily alter the body of that email; they say the same of code-then-execute. On an operating-system assistant, a plan limited to the find and mv commands could still be talked into copying sensitive files to a public network drive, and in their email-assistant case recipients and bodies "could still be influenced by untrusted data." The protection also stops at the user's own prompt. The authors say the first five patterns still allow injections there, and the sixth, context-minimization, is meant to prevent some of them by removing the user's prompt from the model's context once it has shaped the action.
An agent's plan fixed before it reads untrusted data still leaves each call's arguments open to injectionBI-FIND-2025-0272
Cited in 2 briefings
Every briefing that discusses this term, newest first.
- A guide to securing agent plans has the planner choose each step's tools and leaves out agent identity
Paper · · arXiv · Ron F. Del Rosario et al. · Sep 10, 2025
- An agent's plan fixed before it reads untrusted data still leaves each call's arguments open to injection
Paper · · arXiv · Luca Beurer-Kellner et al. · Jun 10, 2025