CertSafari
    Snowflake SnowPro Advanced: Security Engineer (SEA-C01)· Lessons

    Domain 4 · Lesson 15/21

    Threat modeling a Snowflake account: assets, entry and exit points, and mitigations

    Perform threat modeling, identification, and analyses.

    21 min read
    4.5% of exam
    14 sources
    Published 20 Oct 2026
    Docs as of 18 Oct 2026

    What you will be able to do

    • Build a catalog of critical assets from classification tags and read it back from TAG_REFERENCES
    • Enumerate every data entry and exit point with SHOW STAGES, SHOW INTEGRATIONS and SHOW SHARES, and know when those commands under-report
    • Apply STRIDE to Snowflake-specific scenarios such as ACCOUNTADMIN sprawl, compromised service credentials and third-party connections
    • Identify threats from data sharing configurations, over-privileged roles and users, compromised service account credentials, and third-party connections and packages, and name where each one leaves evidence
    • Pair each threat with a control and verify the mitigation through Trust Center violations and detections

    Key concept

    Account threat model — A threat model for a Snowflake account is a written map. It lists the data worth protecting and every path by which data enters or leaves the account, and it pairs each path with the threats against it and the control that closes it. If a path is missing from the map, nothing in the model protects it.

    1.Cataloging critical assets from classification tags

    Threat modeling starts with knowing what an attacker would want. In a Snowflake account that usually means the columns that hold personal data, and the roles and credentials that can reach them. Before you can rank threats you need an inventory of those columns, and in an inherited account with hundreds of undocumented tables you cannot build it by hand.

    Automatic classification builds the inventory for you. In one sentence: a classification profile set on a database runs sensitive-data classification on a schedule and tags what it finds (the mechanics are in Sensitive data classification (automatic, custom and SYSTEM$CLASSIFY)). With 'auto_tag': true in the profile, every column judged sensitive gets two system tags. SNOWFLAKE.CORE.SEMANTIC_CATEGORY records what kind of data the column holds, such as a name or an email address. SNOWFLAKE.CORE.PRIVACY_CATEGORY records how sensitive it is: IDENTIFIER, QUASI_IDENTIFIER or SENSITIVE. The privacy category gives you a ranking for free. IDENTIFIER columns are the top tier of critical assets.

    Classification looks at the data, not at column names. A query against INFORMATION_SCHEMA.COLUMNS with filters like '%SSN%' only finds columns whose names happen to match, and anything named c7 or cust_ref slips through. In Snowflake's own tutorial, a column called ACCOUNT_NUMBER was not classified as sensitive, while EMAIL came back with semantic category EMAIL and privacy category IDENTIFIER. Once the profile is set, Snowflake classifies the existing data and also classifies new data as it is added, so the catalog stays current without a second project.

    Two other sources of data look like they might do this job, and neither does. Trust Center scanners check account configuration, such as MFA enforcement or admin role counts. They do not report which tables contain personal data. ACCESS_HISTORY records which columns each query read or wrote, so it shows what is popular. Plenty of heavily queried columns are not sensitive, and a sensitive column nobody queried last year is still a critical asset.

    The SNOWFLAKE.ACCOUNT_USAGE.TAG_REFERENCES view is where you read the catalog back. Each row links one tag to one object and has TAG_NAME, TAG_VALUE, DOMAIN (COLUMN for column-level assignments), COLUMN_NAME, and OBJECT_NAME, which holds the parent table name when the tag is on a column. APPLY_METHOD = CLASSIFIED marks tags that classification applied automatically, as opposed to MANUAL or PROPAGATED ones. Three limits matter for an inventory. The view records only direct tag-to-object links, so inherited tags do not appear. Columns that have been deleted do not appear. And latency can reach two hours. The TAGS view, by contrast, holds tag definitions, and you join to it only for details like TAG_DATABASE_ID.

    Active tag associations across the account. Add a filter on tag_name = 'PRIVACY_CATEGORY' to get the ranked asset list.sql
    select tag_name, tag_value, domain, object_id from snowflake.account_usage.tag_references where object_deleted is null order by tag_name, domain, object_id;

    Checkpoint 1 of 9· Check yourself

    You inherit a Snowflake account with hundreds of undocumented tables and need a list of critical assets (the columns holding personal data). A colleague suggests building it from the ACCOUNT_USAGE.ACCESS_HISTORY view. What is wrong with that plan?

    Checkpoint 2 of 9· Exam question

    A security team inherits an account with roughly 400 tables spread across 30 schemas and must catalog every column that holds personal data so it can be ranked as a critical asset. Nobody documented the data. Which approach builds this inventory with the least manual effort?

    Checkpoint 3 of 9· Exam question

    Which ACCOUNT_USAGE view should an analyst query to list every table column that currently carries the PRIVACY_CATEGORY tag, together with the tag value, when ranking assets by sensitivity?

    Sources1234

    2.Enumerating entry and exit points

    After the assets come the doors. Any object that connects the account to something outside it is an entry point, an exit point, or both. In Snowflake those objects are stages, integrations and shares, and each has a SHOW command that lists them. Warehouses, resource monitors and tag definitions control compute, credits and metadata. None of them moves data across the account boundary, so they do not belong in this inventory.

    Commands that enumerate data paths, and which output columns describe each path
    CommandColumns to recordPath it reveals
    SHOW STAGES IN ACCOUNTurl, type, storage_integration, has_credentials, owner_role_typeExternal cloud locations used to load data in and unload data out
    SHOW INTEGRATIONSname and type (API, CATALOG, EXTERNAL ACCESS, NOTIFICATION, SECURITY, STORAGE)Connections to outside systems: buckets, APIs, identity providers, egress destinations
    SHOW SHARESkind (INBOUND or OUTBOUND), to, database_nameData offered to other accounts, or consumed from them

    Stages carry both inbound loads and outbound unloads. An external stage has a url (blank for internal stages) and usually a storage_integration that grants it bucket access. In the threat model, each external stage is a COPY destination, and you need to check which account parameters close it, as described in Account parameters restricting unload destinations. An owner_role_type of APPLICATION means a Snowflake Native App owns the stage, so record it under third-party paths.

    SHOW INTEGRATIONS can be narrowed to one integration typesql
    SHOW [ { API | CATALOG | EXTERNAL ACCESS | NOTIFICATION | SECURITY | STORAGE } ] INTEGRATIONS [ LIKE '<pattern>' ]

    Integrations are the account's named connections to the outside world. Each external access integration is an outbound path from code running in Snowflake to a network destination, so it goes in the exit column (how to configure one is covered in External access integrations). Security integrations are inbound: they define how OAuth clients and identity providers get in.

    Shares move data without copying it, which makes them easy to overlook. An OUTBOUND share means your account is sharing data with other accounts, and every object added to it is visible to consumers immediately. Two output details catch people out. The to column displays at most three accounts, so you need TARGET_ACCOUNTS in the SHARES view for the full list. And SHOW GRANTS OF SHARE lists only accounts that have already created a database from the share. Reviewing listings and reader accounts as exits is covered in Shares, listings and reader accounts.

    Inventories under-report silently. SHOW returns only objects on which your current role holds at least one privilege, unless the role has MANAGE GRANTS. By default only ACCOUNTADMIN and SECURITYADMIN have it, and it lets the holder see every object. Omitting IN ACCOUNT while a database is in use limits the output to that database. The Account Usage STAGES view ignores MANAGE GRANTS and can lag by up to two hours. One entry point appears in none of these lists: external engines such as Spark or Trino that use the Horizon Iceberg REST Catalog. Their activity shows up in ACCESS_HISTORY with event_source = 'horizon_irc'.

    Checkpoint 4 of 9· Check yourself

    To list the account's external load and unload paths, an engineer runs SHOW STAGES IN ACCOUNT while using the SYSADMIN role and records the 12 rows returned as the complete stage inventory. What should the threat model say about that list?

    Checkpoint 5 of 9· Exam question

    A newly appointed engineer must document every path through which data can enter or leave an account. Which THREE commands, run across the account, help enumerate those paths?(Select 3)

    0 of 3 selected

    Sources567894

    3.Applying STRIDE to Snowflake scenarios

    STRIDE gives each asset and path a checklist of six questions: can someone Spoof an identity, Tamper with data, Repudiate an action, Disclose information, Deny service, or Elevate privilege? The method is generic. The value comes from answering each question with a specific Snowflake object, signal and control. The table below maps five of the categories to Snowflake scenarios. Denial of service is mostly a cost and availability question and is better handled in risk assessment.

    Threat categories mapped to Snowflake threat scenarios and the evidence for them
    Threat categorySnowflake scenarioWhere it shows up
    SpoofingStolen service-user key or password used from an unknown networkTrust Center detection: login from an unrecognized IP address; risky-user findings
    TamperingUnauthorized INSERT, UPDATE, DELETE or DDL on a critical tableACCESS_HISTORY objects_modified and object_modified_by_ddl
    RepudiationA user denies running a write against regulated dataACCESS_HISTORY ties user_name and query_id to the objects touched
    Information disclosureBulk unload to an external stage, an over-broad outbound share, or egress through an external access integrationTrust Center detection: large data transfer to an external stage; the exit-point inventory
    Elevation of privilegeACCOUNTADMIN granted to many users, or roles with grants far beyond their functionTrust Center scanners for over-privileged roles and for ACCOUNTADMIN/SECURITYADMIN user counts

    Elevation of privilege deserves extra attention in Snowflake, because the scenario often comes from the account's own defaults rather than from an attacker. The first user provisioned gets ACCOUNTADMIN, and when nobody revisits that, admin grants spread. When you model this threat, count ACCOUNTADMIN holders and look for roles whose grants exceed the business function they are meant to serve. The structure you are measuring against is set out in RBAC role hierarchy and least privilege (system roles, functional vs access roles). Keep in mind that ACCOUNTADMIN is not a superuser. It can only act on objects that it, or a role below it in the hierarchy, has privileges on. Objects owned by a custom role that was never granted into the SYSADMIN tree are out of its reach, including for cleanup.

    Compromised service credentials are a spoofing threat with an elevation multiplier. An automated script that runs as ACCOUNTADMIN turns one leaked key into full control of the account. Model each service user as an asset-and-path pair: which role it holds, which networks it can authenticate from, and which exits that role can reach. The authentication methods are covered in Key-pair authentication, programmatic access tokens and service users. In the model, what matters is that the credential's blast radius is set by its role and its network policy.

    Third-party code is a supply-chain path. A Native App can own stages and other objects, and Trust Center can flag client versions such as drivers that have known vulnerabilities or have reached end of life. Review installed apps the way you review a vendor, using the controls in Native App security review (NAAAPS) and secure app code.

    Checkpoint 6 of 9· Check yourself

    A nightly ETL script authenticates as a service user whose only role is ACCOUNTADMIN. In the threat model, this pairing is flagged as spoofing with an elevation multiplier. Which mitigation most directly shrinks the blast radius if the script's key leaks?

    Sources101145

    4.Threat classes specific to Snowflake

    Threat modeling methodologies such as STRIDE are generic, so a model is only as good as the Snowflake-specific threats fed into it. Four classes recur in Snowflake accounts, and each one leaves evidence in a different place:

    • Data sharing configurations. A share is controlled completely by the provider account, and new objects added to it become available to every consumer immediately. The threats are an object added to a share by mistake, an account added that should not be, and sensitive data offered to a non-Business Critical account, which Snowflake supports but discourages. Review the objects with SHOW GRANTS TO SHARE and the full consumer list with TARGET_ACCOUNTS.
    • Over-privileged roles and users. Roles accumulate grants beyond their function, and users keep access they no longer use. Trust Center scanners find over-privileged roles, check that the number of ACCOUNTADMIN and SECURITYADMIN users is limited, and find users who have not logged in for 90 days. A dormant user holding a powerful role is a path nobody is watching.
    • Compromised service account credentials. A key or password for a non-human user works from anywhere its network policy allows and can do whatever its role can do. Record each service user's role and network policy, and watch for Trust Center risky-user findings and logins from unrecognized IP addresses.
    • Vulnerabilities in third-party connections and packages. Drivers and clients in use can have known vulnerabilities or be past end of life, which Trust Center reports as a detection. Each integration listed by SHOW INTEGRATIONS is a connection to an outside system, so check its enabled value and whether it is still needed, and list packages policies with SHOW PACKAGES POLICIES.

    Write each threat as one line in the model: the asset or path, what the attacker does, the control, and the signal that proves it happened. A share threat reads: outbound share SALES_S, a consumer account added in error, access revoked and grants reviewed, a changed TARGET_ACCOUNTS list as the signal. A line that cannot name its signal is a threat you would only learn about from the consumer.

    Checkpoint 7 of 9· Check yourself

    A provider adds a table with IDENTIFIER columns to an outbound share that consumer accounts have already imported. When can those consumers query the table?

    Sources9121013

    5.Violations and detections as threat evidence

    Trust Center checks parts of the threat model continuously. Its scanners produce two kinds of findings, and they behave differently enough that mixing them up leads to wrong actions.

    A violation is a configuration that persists over time. Examples are users without MFA, too many ACCOUNTADMIN holders, and over-privileged roles. Each one corresponds to a threat in the model whose control is missing. You fix it by changing the configuration. The violation stays on the Violations tab until the next scheduled run of its scanner package, or until you run the package on demand. Marking a violation Resolved records a justification and suppresses email notifications, but the configuration underneath is unchanged.

    A detection is a single event: a login from an unrecognized IP address, a large transfer to an external stage, or an unusually high task error rate. Because the event has already happened, there is nothing to remediate directly, and you cannot manage a detection's lifecycle. What you can do is investigate it, and if it turns out to be real, tighten the control on the path it used. In threat-model terms, a detection is evidence that a threat you listed is active.

    Checkpoint 8 of 9· Check yourself

    The Trust Center reports that a large amount of data was transferred to an external stage last night. The security engineer wants to close this out properly. Which response fits this kind of finding?

    Sources10

    6.Implementing and verifying the mitigations

    A mitigation counts only if it is attached to a specific threat on a specific path. For elevation of privilege through ACCOUNTADMIN, the controls are concrete:

    • Grant ACCOUNTADMIN to a limited number of people, and to at least two, so that a lost password does not trigger a recovery process that can take up to two business days.
    • Require MFA for every ACCOUNTADMIN user.
    • Do not make ACCOUNTADMIN anyone's default role, so that users have to switch to it deliberately.
    • Run automated scripts under a role in the SYSADMIN hierarchy, not ACCOUNTADMIN.

    Network policies are the control on each inbound path, and the threat model should record which policy guards which entry point. In one sentence: a policy allows or blocks request origins through network rules and applies at account, user or security-integration level (see Network policies: allowed/blocked lists and where they apply). For a service user, the strongest mitigation against credential theft is a dedicated network rule that allows only that workload's addresses, applied to that user. Snowflake recommends exactly this kind of rule for highly privileged users and service accounts. A stolen key used from anywhere else is rejected even though it is valid.

    Two mistakes leave a documented control doing nothing. First, a policy has no effect until it is activated on the account, a user or an integration, so record where it is applied, not just that it exists. Second, precedence runs from most specific to least: an integration-level policy overrides a user-level one, which overrides the account-level one. A permissive user-level policy therefore weakens a strict account policy for that user.

    Verification closes the loop. For configuration threats, run the relevant Trust Center scanner package on demand after the change and confirm the violation no longer appears. For exit paths, rerun the SHOW inventory as a role with MANAGE GRANTS and compare it against the model. For data threats, ACCESS_HISTORY should show only the expected users reading and writing your IDENTIFIER-tier columns, because each record links the user, the query, the object and the column.

    Checkpoint 9 of 9· Match them up

    Match each network policy state or placement to its effect when the threat model records it as the control on an entry point.

    Tap a term, then the definition that fits it.

    Sources1114104

    Practise it for real

    Produce a first-pass asset and data-path inventory for one account

    1. 1.As a role with MANAGE GRANTS, query snowflake.account_usage.tag_references where object_deleted is null and filter on tag_name = 'PRIVACY_CATEGORY'.

      Why: Classification tags rank columns by sensitivity, which gives you the critical-asset tier list.

      You should see: Rows with DOMAIN = COLUMN, OBJECT_NAME holding the parent table, and TAG_VALUE such as IDENTIFIER.

    2. 2.Run SHOW STAGES IN ACCOUNT and record url, type, storage_integration and owner_role_type for each external stage.

      Why: External stages are load and unload paths to cloud storage.

      You should see: Internal stages show a blank url. External stages show a cloud URL and usually a storage integration.

    3. 3.Run SHOW INTEGRATIONS, then SHOW EXTERNAL ACCESS INTEGRATIONS.

      Why: Integrations are the account's named connections to outside systems, and external access integrations are egress paths.

      You should see: A list of integrations by type, with external access integrations listed separately.

    4. 4.Run SHOW SHARES and, for each OUTBOUND share, query TARGET_ACCOUNTS in snowflake.account_usage.shares.

      Why: The to column shows at most three accounts, so the view gives the full consumer list.

      You should see: INBOUND and OUTBOUND rows, plus the complete set of target accounts for each outbound share.

    Stuck? Get a nudge

    If a stage or share you know about is missing from the output, check which role you are using before you conclude it does not exist.

    Sources

    (14)

    Every claim above is drawn from one of these pages, quoted as it was written on the date shown.

    1. 1.
      “Classification tags will be automatically set on columns identified as containing sensitive data ('auto_tag': true).”
      ↩︎ Cataloging critical assets from classification tags
      “Snowflake starts classifying the existing data and classifies new data when it is added to the database.”
      ↩︎ Cataloging critical assets from classification tags
      “The ACCOUNT_NUMBER column was not classified as sensitive, so it wasn’t assigned classification tags.”
      ↩︎ Cataloging critical assets from classification tags
    2. 2.
      “A privacy category identifies the sensitivity of a personal attribute. It can be IDENTIFIER, QUASI_IDENTIFIER, or SENSITIVE”
      ↩︎ Cataloging critical assets from classification tags
      “It’s critical to know where your sensitive data resides and if it’s adequately protected.”
      ↩︎ Key concept
    3. 3.
      “CLASSIFIED: The tag was automatically applied to a column that was classified as containing sensitive data.”
      ↩︎ Cataloging critical assets from classification tags
      “The view does not contain information about columns that have been deleted.”
      ↩︎ Cataloging critical assets from classification tags
      “This view only records the direct relationship between the object and the tag. Tag inheritance is not included in this view.”
      ↩︎ Exam trap 1
    4. 4.
      “provide insights on popular and frequently accessed tables and columns”
      ↩︎ Cataloging critical assets from classification tags
      “You can identify IRC-sourced records by filtering on event_source = 'horizon_irc'”
      ↩︎ Enumerating entry and exit points
      “The object_modified_by_ddl column records the DDL operation on a database, schema, table, view, and column.”
      ↩︎ Applying STRIDE to Snowflake scenarios
      “there is a direct link between the user (i.e. query operator), the query, the table or view, the column, and the data.”
      ↩︎ Implementing and verifying the mitigations
    5. 5.
      “URL for the external stage; blank for an internal stage.”
      ↩︎ Enumerating entry and exit points
      “The command only returns objects for which the current user’s current role has been granted at least one access privilege.”
      ↩︎ Enumerating entry and exit points
      “Using SHOW commands without an IN clause in a database context can result in fewer than expected results.”
      ↩︎ Enumerating entry and exit points
      “If a Snowflake Native App owns the object, the value is APPLICATION.”
      ↩︎ Applying STRIDE to Snowflake scenarios
      “The MANAGE GRANTS access privilege implicitly allows its holder to see every object in the account.”
      ↩︎ Exam trap 2
    6. 6.
      “The view does not recognize the MANAGE GRANTS privilege and consequently may show less information compared to a SHOW command”
      ↩︎ Enumerating entry and exit points
    7. 8.
      “OUTBOUND indicates that your account is sharing data with other accounts and this share was created in your account.”
      ↩︎ Enumerating entry and exit points
      “SHOW GRANTS OF SHARE lists only accounts that have created a database from the share.”
      ↩︎ Enumerating entry and exit points
    8. 9.
      “With Secure Data Sharing, no actual data is copied or transferred between accounts.”
      ↩︎ Enumerating entry and exit points
      “New objects added to a share become immediately available to all consumers, providing real-time access to imported data.”
      ↩︎ Threat classes specific to Snowflake
      “New objects added to a share become immediately available to all consumers, providing real-time access to imported data.”
      ↩︎ Exam trap 3
    9. 10.
      “Login events originated from an unrecognized IP address.”
      ↩︎ Applying STRIDE to Snowflake scenarios
      “Ensure the number of users with the ACCOUNTADMIN and SECURITYADMIN system roles is limited.”
      ↩︎ Applying STRIDE to Snowflake scenarios
      “Find over-privileged roles.”
      ↩︎ Threat classes specific to Snowflake
      “Find users who have not logged in for 90 days.”
      ↩︎ Threat classes specific to Snowflake
      “Some client versions (for example, drivers) currently in use may have known vulnerabilities or have reached their end-of-life status.”
      ↩︎ Threat classes specific to Snowflake
      “A violation persists over time and represents a configuration that doesn’t conform with a scanner’s requirements.”
      ↩︎ Violations and detections as threat evidence
      “A detection occurs one time and represents a unique event.”
      ↩︎ Violations and detections as threat evidence
      “Email notifications are suppressed for resolved violations.”
      ↩︎ Violations and detections as threat evidence
      “Currently, you can’t view detection findings at the organization level.”
      ↩︎ Violations and detections as threat evidence
      “After you remediate the issue, the Trust Center no longer reports the violation.”
      ↩︎ Implementing and verifying the mitigations
      “Because the event is unique and happened in the past, direct remediation of a detection isn’t possible.”
      ↩︎ Exam trap 4
      “For example, a scanner reports a violation if some users haven’t configured multi-factor authentication (MFA).”
      ↩︎ Prediction
      “Because the event is unique and happened in the past, direct remediation of a detection isn’t possible.”
      ↩︎ Checkpoint
    10. 11.
      “Note that ACCOUNTADMIN is not a superuser role.”
      ↩︎ Applying STRIDE to Snowflake scenarios
      “By default, not even the ACCOUNTADMIN role can modify or drop objects created by a custom role.”
      ↩︎ Applying STRIDE to Snowflake scenarios
      “Assign this role only to a select/limited number of people in your organization.”
      ↩︎ Implementing and verifying the mitigations
      “Snowflake recommends using a role other than ACCOUNTADMIN for automated scripts.”
      ↩︎ Implementing and verifying the mitigations
      “The security administrator (SECURITYADMIN system-defined) role includes the global MANAGE GRANTS privilege to grant or revoke privileges on objects in the account.”
      ↩︎ Checkpoint
      “Snowflake recommends using a role other than ACCOUNTADMIN for automated scripts.”
      ↩︎ Checkpoint
    11. 12.
      “Objects that you add to a share are instantly available to the consumer accounts that have created databases from the share.”
      ↩︎ Threat classes specific to Snowflake
      “Snowflake supports sharing sensitive data with non-Business Critical accounts (disabled by default), but does not encourage doing so.”
      ↩︎ Threat classes specific to Snowflake
    12. 13.
      “SHOW PACKAGES POLICIES | Lists packages policy information.”
      ↩︎ Threat classes specific to Snowflake
    13. 14.
      “Creating a network rule whose purpose is to allow access for a special population, such as highly privileged users and service account users.”
      ↩︎ Implementing and verifying the mitigations
      “Network policies applied to a security integration are the most specific network policies.”
      ↩︎ Implementing and verifying the mitigations
      “A network policy doesn’t restrict network traffic until it is activated.”
      ↩︎ Exam trap 5
      “Network policies applied to a security integration are the most specific network policies. They override both accounts and users.”
      ↩︎ Checkpoint

    Ready to test yourself?

    Practise the 16 questions on this subdomain.

    Spotted a mistake, or was something unclear? Tell us.