Snowflake AI Governance & Security

AI adds attack paths that ordinary data platforms lack. You study prompt injection, data exposure through prompts and outputs, over-broad access, unsafe tool use and runaway cost, and relate them to a published list of language model risks.

• Direct and indirect prompt injection
• Sensitive data exposed through prompts and outputs
• Over-broad access granted to AI services
• Unsafe tool use and excessive autonomy
• Cost abuse and runaway usage
• Mapping risks to a published LLM risk list

You build A risk list for one Cortex AI deployment, with the likely attack path for each risk.

A review turns worry into findings. You inspect an account’s AI settings, model availability, roles and grants against a baseline, separate findings that matter from noise, and write them up so an owner can act.

• What to check in an account’s AI settings
• Model availability and allowlists
• Roles and grants on AI features
• Reading a configuration against a baseline
• Findings that matter and findings that do not
• Writing a configuration review

You build A configuration review of a sample account with findings and recommended changes.

Who may call a model, build a search service or query through an analyst service must be decided on purpose. You grant least access to developers, operators and end users, and test the result from each persona’s side.

• Roles that allow calling Cortex functions
• Least access needed to use a model
• Controlling search and analyst services
• Separating developers, operators and users
• Service roles for AI applications
• Testing access as each persona

You build A role design that gives three personas the least access to Cortex AI features.

Controls around a model matter as much as the model. You configure filtering of inputs and outputs, restrict which models and tools may be used, set limits on rate and cost, and test the guardrails with hostile prompts.

• Input and output filtering
• Safety filtering with Cortex Guard
• System prompts as policy and their limits
• Allowlists for models and tools
• Rate and cost limits
• Testing guardrails with adversarial prompts

You build A guardrail configuration tested against a set of adversarial prompts.

Governance rules do not pause when a model is involved. You classify data before it reaches a model, check that masking and row access still apply, and tag, retain and trace what AI generates.

• Classifying data before it reaches a model
• Masking and row access in AI queries
• Tags and lineage for AI-generated data
• Retention of prompts and results
• Personal data inside prompts
• Proving governance holds with a test

You build A test that shows masked data stays masked through an AI workflow.

Where a request is processed can be a legal question. You work out where inference runs, what cross-region settings change, and how to choose regions and models so that residency obligations are met and documented.

• Residency obligations in plain terms
• Where inference runs and where data goes
• Cross-region inference settings
• Choosing regions and models to meet an obligation
• Documenting a residency decision

You build A residency assessment for one deployment saying where each request is processed.

AI can start work that no person typed. You check that governance and protection hold for queries, tasks and tools launched by AI services, and that execution identities cannot be used to slip past a policy.

• Workloads started by AI services
• Whose privileges a triggered query uses
• Policies that apply to AI-issued queries
• Limiting what an agent may run
• Logging and attributing AI-triggered activity

You build A review of an AI-triggered workflow confirming policy applies at every step.

Evidence lives in the usage views. You query them to see who used which AI features, with which models and at what cost, spot unusual volumes, and join AI usage to query history for the full picture.

• Usage and history views for AI features
• Attributing AI spend to users and roles
• Finding unusual model or volume use
• Joining query history with AI usage
• Retention and delay in audit data
• Saving evidence queries

You build A set of evidence queries reporting AI use by user, role and model.

Posture is a trend, not a snapshot. You read Trust Center findings and severity, relate them to AI workloads, decide what to fix first by exposure, and track how posture changes over time.

• What the Trust Center scans for
• Reading findings and severity
• Relating findings to AI workloads
• Deciding what to fix first
• Tracking posture over time

You build A posture report that ties Trust Center findings to specific AI workloads.

You review one deployment from risk assessment to evidence: configuration, access, guardrails, governance, residency and posture. Then you write the opinion a security committee would need to approve or reject it.

• Scoping the deployment and its data
• Assessing risk, access and guardrails
• Testing governance and residency claims
• Gathering usage and posture evidence
• Writing the review opinion

You build A governance review of a Cortex AI deployment with evidence and a recommendation.