Cortex AI governance
Cortex AI is open by default. Governing it takes four settings.
Access, model choice, inference region and audit, and how they interact.
Kevin Keller · kevinkeller.org
Why it matters
The defaults favour access, not control
SNOWFLAKE.CORTEX_USERis granted toPUBLIC: every user gets every Cortex feature- Inference can route to US regions unless you restrict it
- Cortex Analyst can silently fall back to weaker models
- There is no log field showing which region ran a request
Control 1: feature access
Revoke from PUBLIC, then grant on purpose
| Database role | Grants |
|---|---|
CORTEX_USER | All Cortex AI features |
CORTEX_ANALYST_USER | Cortex Analyst only |
CORTEX_ADMIN | Admin capabilities |
Per-user budgets: revoke the AI role when a user exceeds their limit, reinstate with a task.
Control 2: model access
Model RBAC first, then the allowlist
- If a model application role grants access: allow
- If
CORTEX_MODELS_ALLOWLISTlists the model, or is'All': allow - Otherwise: fail with a model-authorisation error
Restrictions apply across all Cortex features, not one at a time.
Breaking change
BCR-2220 extends model RBAC to every AI function
- From the 2026_02 bundle,
AI_TRANSLATE,AI_SENTIMENT,AI_EXTRACTand others enforce model RBAC - They use internal aliases such as
arctic-translate - With a customised allowlist, those calls fail unless the aliases are added
Control 3: inference region
CORTEX_ENABLED_CROSS_REGION decides where prompts can go
| Value | Behaviour |
|---|---|
DISABLED | Home region only |
AWS_EU | AWS EU regions only |
ANY_REGION | Any supported region |
Only ACCOUNTADMIN can set it. Prompts and outputs are never stored during cross-region inference.
The hidden risk
The three controls multiply together
EU region limits, allowlists and model RBAC can push Cortex Analyst down its fallback chain, or block it entirely. That is a silent quality risk, not always a visible error.
Model hosting
Native models avoid cross-region routing
| Family | Compute | EU-native |
|---|---|---|
| Llama, Arctic, Mistral, DeepSeek | Snowflake GPUs | All EU regions |
| Claude | AWS Bedrock | Frankfurt, Ireland |
| GPT, o-series | Azure AI Foundry | West Europe |
All run inside Snowflake's governance perimeter.
Control 4: audit
Each surface has its own view
- AI functions:
CORTEX_AISQL_USAGE_HISTORY, 365 days - Real time:
CORTEX_AI_FUNCTIONS_USAGE_HISTORYtracks in-flight queries and roles - Analyst:
CORTEX_ANALYST_USAGE_HISTORY, per-message billing - Agents: immutable traces in
AI_OBSERVABILITY_EVENTS
The audit gap
Log which model actually answered
Usage views don't record the resolved model for Cortex Analyst.
Log response_metadata.model_names with each request_id from the REST API. Alert if a frontier model appears on an account meant to use native models only.
Takeaway
Revoke the open default, pin models and regions, and log which model really ran.
Kevin Keller · kevinkeller.org