Federal Zero Trust
One control plane for data and AI inference
Mapping Snowflake and Cortex to NIST 800-207, CISA ZTMM and OMB M-22-09.
Kevin Keller · kevinkeller.org
Why it matters
The mandate predates production AI
OMB M-22-09 requires Zero Trust. It was written before LLM workloads were real. Agencies now have to extend it to model calls, agent tool calls and retrieval, without a second governance stack.
The problem
AI bolted on as a separate layer breaks Zero Trust
- Shadow AI: data sent to external model APIs, bypassing classification
- No audit trail: who invoked which model, with what data?
- Policy divergence: PII masked in SQL, raw in the AI pipeline
- Network blind spots: AI on public endpoints while data needs PrivateLink
The fix
Run inference inside the data platform
The same identity, network, RBAC, masking, tagging and audit govern both. Horizon governs data and AI.
Policy before the model
The model never sees data the role can't access
If a masking policy hides SSNs for ANALYST_ROLE, a Cortex call running as that role receives masked input.
Enforcement happens before the data reaches the model, not after.
Agents and MCP
Every tool call is a governed query
- Cortex Agents, Search and managed MCP act as a Snowflake role
- RBAC, row access, masking and tags apply to every call
- External agents connect via OAuth or key-pair, under RBAC
- Every MCP tool call lands in
QUERY_HISTORYwith lineage
The AI mapping
Zero Trust questions, one answer each
| Question | Approach |
|---|---|
| Who can invoke models? | RBAC on Cortex functions |
| What can the model see? | Row access and masking first |
| Where does inference run? | Inside Snowflake's boundary |
| Is AI usage audited? | QUERY_HISTORY and ACCESS_HISTORY |
The five CISA pillars
Foundation controls that make governed AI possible
- Identity: SSO, MFA, SCIM, key-pair auth, session policies
- Devices: IdP conditional access before the token is issued
- Networks: PrivateLink, private-only access, per-user policies
- Applications: scoped OAuth, role and network policy per workload
- Data: tags, masking, row access, Tri-Secret Secure
Federal context
Dedicated government infrastructure
| Certification | Scope |
|---|---|
| FedRAMP High | AWS GovCloud US-East and West |
| DISA IL-5 | DoD workloads, CUI |
| FIPS 140-2 | Validated cryptographic modules |
| ITAR | U.S. data residency and person access |
Getting started
Seven steps, in order
- Assess the current state
- Identity first: SSO, MFA, SCIM, no password service accounts
- Lock the network
- Classify and protect data
- Give Cortex applications dedicated roles
- Export
ACCOUNT_USAGEto your SIEM - Validate continuously with Trust Center
The takeaway
One governance layer for data and AI: no shadow AI, no blind spots.
Kevin Keller · kevinkeller.org