Protect your enterprise AI in Snowflake with Cyera

Nov 17, 2026
Share

“The goal is not just to discover data or classify the data. It’s to really help Snowflake users and organizations understand data well enough to make decisions and take actions.” - Gal Pollak, Senior Product Manager at Cyera

A simple question like 'who are our highest-risk patients?' can trigger more risk than one would expect. When asked through a Cortex agent in Snowflake, the AI pulls data directly from the source. If those columns include HIPAA-protected health information, the agent does not recognize the sensitivity. There is no warning. The data is exposed without anyone noticing.

This is the gap that Snowflake data governance for AI aims to close, and it was the subject of a webinar Cyera held with Snowflake, where Gal Pollak, Senior Product Manager at Cyera, who leads the joint integration work with Snowflake, and Navnit Shukla, Principal AI Architect at Snowflake, examined the topic. Gal’s starting point:ֿ

“If your network is the walls of a building, for example, AI can move straight into the filing cabinets. So security can start closer to what matters most, which is the data itself.”

In this scenario, all controls worked as configured. The agent’s role allowed access to the columns, and no masking policy applied, so the data was returned without redaction. The column was never identified as sensitive, so no one had protected it. That is the gap Gal's points at: an agent acted on data nobody had understood well enough to control.

A masking policy only protects the columns someone has already found

Think of roles as the gatekeepers for who can access a table - masking policies go one step further, controlling what each role can see in a column. That means sensitive data can be hidden or scrambled for some users, while remaining visible to others.

A masking policy protects a column once it has been assigned to it, directly or through a tag (a label set on the column or its table). A credit card column that nobody has found, and so nobody has tagged or protected, is readable by every role that can query the table, and by every agent running under one of those roles (agents query Snowflake through a role, the same way people do).

Gal explained: “AI tools, AI agents, have very wide access that sometimes is almost invisible,” and they use it exhaustively, whereas people barely touch theirs. On timing, she said, “AI projects do not wait for slow security programs. If the business is moving quickly, security has to keep up.” Gal described the agent-specific risk in four parts:

1. Agent sprawl 2. Sensitive data exposure: 3. Overprivileged by default: 4. High-risk and irreversible actions
Agents deployed across teams, tools and vendors, with no single owner or audit trail. PII, financials and intellectual property flowing in and out of AI with no controls. Agents operate as super users, accounts with access to the most sensitive data. Agents can execute, write, delete and send operations with no oversight.
All four return to the same question: what data can this agent reach, and should it? That is why the work starts with the data rather than with the agent.

Start by understanding your sensitive data. Identify where it lives, which agents can access it, and whether that access is necessary. Before you write a masking policy, confirm five details:

  • Which columns hold sensitive data?
  • Which roles and agents can reach them?
  • Which of those columns carry a masking policy?
  • What job was each Cortex agent given?
  • What data does it touch in practice?

Snowflake data governance for AI: four steps every practitioner should master, from first column to last agent

The five questions tell you what you need to know, and the four steps below are how you get the answers and act on them.

Gal put it simply: to scale AI agents safely, you need a foundation that knows your data, who can access it, and what agents are doing with it. That foundation is built on five capabilities: discover, govern, protect, validate, and enable. Enable is what you get once the first four are working. For Snowflake teams, here’s how those first four play out in practice:

  1. Discover: Map sensitive data across Snowflake, down to each column, and classify it: personal and health data (PII and PHI), financial records and the data types specific to your business. Record which users, roles and AI agents can reach each column, and keep that current as access changes.
  2. Govern: Turn each classification into a Snowflake tag and attach masking policies to the tags, so a credit card column found later picks up the policy once it is tagged.
  3. Protect: Fix the problem where Cyera found it. When Cyera raises an issue for credit card numbers in plain text (unmasked), applying the tag is the remediation, because the policy follows the tag.
  4. Validate: Inventory every Cortex agent, map its full data reach (the tables and columns it can query), and check what it touches against its intended purpose.

Cyera for Snowflake covers those four steps with 3 pieces. As Gal puts it: “Cyera is the layer that makes AI trustworthy at scale.” DSPM handles discovery and classification as it finds the sensitive columns and classifies them. Cyera’s Snowflake Access Governance turns the classifications into tags, masking policies and tighter roles. Agent Guardian does the validating, for the agents themselves. It builds a list of every Cortex agent running in your Snowflake environment, shows which tables and columns each one can reach, points out where an agent can get to data it has no business reason to touch, and lets you cut that access with one click instead of opening a ticket. As Gal said of Agent Guardian: “This is how you can scale your AI agents on Snowflake without scaling risk.” Her one-line version of the whole thing: “Cyera is the layer that makes AI trustworthy at scale.”

Every step rests on whether the column was classified correctly

During the demo, Gal showed that the four steps are one motion and presented the chain. She opened an issue Cyera had raised for credit card numbers in plain text and started the remediation, which applies a Snowflake tag. If that tag carries a masking policy, every column holding card numbers is masked at once. Apply the same logic to the opening scenario: tag the patient columns, attach a masking policy, and the agent’s query still runs, but any role the policy covers gets masked values instead of the real ones.

That only holds if the classification is right. Gal said 86% of Cyera’s classifications are learned data types: the engine learns the data types unique to each business instead of matching a generic pattern. Cyera has classified over an exabyte of Snowflake data with 95% precision and identified more than 1 trillion sensitive records at risk of exposure.

The clearest proof came from a customer. A UK fintech running on AWS and Snowflake had personal customer data spread across its warehouse, documents, and spreadsheets as it grew, and needed to know where it lived, who was viewing it, and where it flowed. Cyera’s first full scan found around 170 million potential sensitive instances. Through identification and minimization, the team brought that down to 500,000, and stopped personal data flowing out of its core systems into other databases. 

None of this is an argument for holding agents back until governance is perfect. Gal’s position was to adopt AI early and treat understanding the data as the price of doing it securely. Agents won’t wait for governance to catch up, so the governance has to live on the data: find the column, tag it, and let the masking policy follow. That is what understanding data well enough to act looks like in Snowflake.

“The goal is not just to discover data or classify the data. It’s to really help Snowflake users and organizations understand data well enough to make decisions and take actions.” - Gal Pollak, Senior Product Manager at Cyera

Watch the full webinar on demand

Govern the Data Powering AI with Cyera and Snowflake is available on demand, with the demo in full: the issue list Cyera builds for Snowflake, the tag that triggers masking, and Agent Guardian’s view of what each Cortex agent can reach.

Share