Source: Agents Rule of Two (Meta)

Meta (Oct 2025): within one session an agent should hold at most two of untrusted input, sensitive data/systems, and state change or external communication.

Contents

Summary

A design rule for limiting the blast radius of prompt injection until models can reliably detect and refuse it. Properties: [A] processes untrustworthy inputs; [B] accesses sensitive systems or private data; [C] changes state or communicates externally. Any session may combine at most two.

Key claims

  • If all three are required, the agent must not run autonomously: add human-in-the-loop approval or another reliable validation, or split the work across fresh sessions.
  • An agent may switch configuration mid-session, e.g. turn off external communication before touching internal systems.
  • Examples: email bot (restrict to trusted senders [BC] or human-reviewed drafts [AB]); travel assistant [AB] with confirmation for bookings; web research assistant [AC] in a sandbox with no session data; internal coder [BC] filtering inputs by author lineage.
  • Not covered: hallucinations, spam, over-privilege generally, lower-impact injection outcomes such as misinformation. It supplements least privilege and defence in depth rather than replacing them; blindly approved confirmations still fail.

Feeds

Prompt injection · LLM security risks

Suggest changes →