Context Trust Firewall
A document, webpage, email or tool result is data. It must not silently become a privileged instruction.
Trust labels
Section titled “Trust labels”SYSTEM_POLICY highest authorityDEVELOPER_POLICY application authorityUSER_INSTRUCTION authorized user intentTRUSTED_DATA verified informationUNTRUSTED_DATA external or retrieved contentGENERATED_DATA model-produced contentNormative target
Section titled “Normative target”Untrusted context must never be permitted to:
- increase its own authority;
- grant tool permissions;
- alter system or compliance policy;
- disable logging;
- change approved destinations;
- request disclosure of protected context.
Planned controls
Section titled “Planned controls”- source-level and span-level taint tracking;
- prompt-injection detection;
- capability firewall for tools;
- output exfiltration checks;
- provenance-preserving transformations;
- signed and explainable deny decisions;
- adversarial conformance fixtures.