Don't Trust a Second LLM: A Better Security Architecture for Local Agents
sunychoudhary · reddit · 2026-07-30
The author questions the popular main model -> guard model -> execution architecture for local agents, arguing that a second LLM guard model fails to create a reliable security boundary. High sensitivity blocks legitimate technical work, while low sensitivity allows indirect prompt injections.
Instead, the author proposes a more defensible architecture:
- Models propose structured actions rather than executing them directly.
- Deterministic policies validate tools, arguments, paths, and data classifications.
- Retrieved documents are treated as untrusted data.
- High-impact actions require explicit human approval.
- The guard model contributes a risk score but does not make the final authorization decision.
More from coding & agent
- ClaudeSkills: 13 End-to-End Automation Skills to Ship Finished Work with Claude Code — tom_doerr · 2026-07-30
- From Coding Tool to Personal OS: A Developer's Headless Agent Setup & Memory Design — PlentyWrongdoer3034 · 2026-07-30
- 4 Types of RAG Explained: From Vanilla Retrieval to Agentic Workflows — PawelHuryn · 2026-07-30
- AI Hero Launches Free AI Skills Course for Engineers — mattpocockuk · 2026-07-30
- Matt Pocock Releases Wayfinder: AI Skill to Plan Complex Projects and Sync Issue Trackers — mattpocockuk · 2026-07-30
- Developer Turns Claude Code into an Automated SEO Manager (Open-Source) — Caitaline_Evars · 2026-07-30