Learn · Security

Two models, one with no tools.

In the dual-LLM pattern, untrusted content is read by a quarantined model that has no tools. It returns structured fields. The privileged model, which can act, never sees the raw text.

Quarantined

Reads everything, can do nothing

The quarantined model turns an email or page into fields that match a schema: sender, intent, amount, dates.

  • No tools
  • JSON schema output
  • Can be a cheaper model

Privileged

Acts, but only on fields

The privileged model sees the fields and references, never the free text, so injected instructions have nowhere to land.

  • Tools behind policy
  • References instead of raw text
  • Typed commands

In Bitfolk

How the Mayor reads

Bitfolk's Mayor uses this pattern for external content, plus taint tracking for anything that still slips through.

  • Quarantined reader for emails, web, proxy responses
  • Taint on runs that read untrusted data
  • Approvals for risky tainted actions

Questions

Does the dual-LLM pattern cost more?

A little. The quarantined reader can use a small, cheap model because it only extracts fields.

Keep exploring

Safe reading, built in

Security is on in every plan.