Skip to content

Security and Trust Boundaries

v9 note: These trust boundaries remain valid. V9 adds materialization, publication and confidentiality concerns in Security-and-Confidentiality-v9.

Model is an untrusted planner

An AI can propose a typed query, projection or mutation request. It is not the final authority for privileged effects.

model proposal
   ↓
schema validation
   ↓
semantic validation
   ↓
authorization / capability / budget gate
   ↓
deterministic executor
   ↓
receipt / result

Prompt/context taint

Retrieved documents, web content, logs, emails, KOS text and tool outputs are data. Context items should retain provenance, trust class, tenant/scope where applicable, sensitivity classification where applicable, and retrieval reason.

Untrusted context cannot redefine system policy, tool authority or approval requirements.

Schema valid is not enough

A valid JSON object can still reference nonexistent identities, violate semantic constraints, exceed authority scope or request unsupported operations. Implementations distinguish at least:

schema reject
semantic reject
authorization / capability reject
state / admission reject

Budgets and least capability

AI/federated query limits live outside the model. Child queries consume parent budget. Components should receive only the capabilities they need: a language resolver does not need canonical write authority; a context compiler does not need mutation authority.

Logs and evidence

Operational logs help observability but are not canonical evidence unless explicitly admitted under an evidence contract.