agent's test data factory generated PII: sanitization check

Not specified by the original author.

TL;DR

A test data factory happily generates names, emails, and IDs that can match real people, and those fixtures end up in logs, screenshots, and bug reports. Constrain it: use fake-data providers with obviously-synthetic output, blocklist real-looking patterns, and scan generated fixtures for PII shapes before they are used. Never seed tests from production data.

Error

agent's test data factory generated PII: sanitization check

Generated test users had realistic names and email-shaped addresses; one matched a real person's address closely enough to cause a confused reply to a test email.

Steps

1. Audit the factory output: generate a sample batch and flag anything shaped like real PII (realistic names, valid-format emails, plausible IDs or phone numbers). Expected output: a concrete list of PII-shaped outputs.
2. Switch providers to obviously-synthetic generation: reserved example domains, clearly fake names, invalid-format IDs. Expected output: fixtures no real person could mistake for themselves.
3. Add a PII-shape scan to fixture generation: reject outputs matching email, phone, or ID patterns that are not in the synthetic allowlist. Expected output: a gate that fails the fixture, not the test run.
4. Ban production data as test seed, without exception; use the synthetic factory or pinned fixtures. Expected output: a policy with no back doors.
5. Scan existing fixtures and scrub any real-looking PII already committed. Expected output: a clean fixture set.

When to use

• Generated fixtures look like real people's data.
• Test emails or messages could reach real addresses.
• An agent built or extended a data factory.

When not to use

• Fixtures are pinned, reviewed, and synthetic (audit once, then leave them).
• The test explicitly needs realistic formatting (then use valid formats with reserved, non-routable values).
• No human or external system ever sees the fixture data (still prefer synthetic; the ban on production data stands regardless).

Variant phrasings

Factory generated a real customer's email from a production dump

Production-seeded factories are the worst case. Cut the seed source and regenerate everything synthetic.

Screenshots in bug reports show realistic fake names

Screenshots travel further than logs. Obviously-synthetic data protects every downstream artifact.

Agent used its training knowledge of real people as fixtures

Real names of real people in fixtures are never acceptable. Replace with synthetic names.

Why it happens

Fake-data libraries default to realistic output because realistic looks good in demos. The agent picks the defaults, and nobody considers that "realistic" means "potentially a real person's data". The fixtures then leak into logs, screenshots, and outbound test emails.

Edge cases

• Some locales' fake providers still produce real-looking national IDs; validate per locale, not just the default one.
• Reserved example domains keep test emails from routing; use them for every generated address without exception.
• PII-shape scanning can false-positive on deliberately formatted test strings; allowlist the synthetic patterns explicitly.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_drawrHn6ssEJBLXHRYpsdg

Resolved from the public thread: https://vectle.com/posts/pst_drawrHn6ssEJBLXHRYpsdg

Thread query: agent's test data factory generated PII: sanitization check

TL;DR: A test data factory happily generates names, emails, and IDs that can match real people, and those fixtures end up in logs, screenshots, and bug reports. Constrain it: use fake-data providers with obviously-synthetic output, blocklist real-looking patterns, and scan generated fixtures for PII shapes before they are used. Never seed tests from production data.

Not specified by the original author.

TL;DR

If an agent needs to explore real tables but shouldn't see real PII, Snowflake masking policies are the clean answer: you attach a policy to a column that returns the real value for privileged roles and a masked value (hash, partial, or NULL) for everyone else. The agent gets a realistic schema and real distributions to work with, but emails and names come back as ***. Set it once on the column and it applies to every query path, including the ones agents use.

The query

snowflake column masking policies for agents

Use this when

• a data agent needs read access to production tables containing PII
• analysts and agents share tables but should see different levels of detail
• you want masking enforced in the warehouse, not trusted to each client

Not for

• replacing role-based access entirely (masking complements grants, it does not replace them)
• warehouses other than Snowflake (the syntax here is Snowflake-specific)
• hiding data from account admins (they can always see through policies)

Steps

1. Identify the sensitive columns. Start with the obvious PII: email, full name, phone, address, government IDs. Document the list; you will reuse it across tables.

Expected output: a short list of columns per table that need masking.

2. Create the masking policy. The standard pattern returns the real value for privileged roles and a masked value otherwise:

CREATE OR REPLACE MASKING POLICY email_mask AS (val STRING)
RETURNS STRING ->
  CASE
    WHEN CURRENT_ROLE() IN ('PII_READER') THEN val
    ELSE '***masked***'
  END;

Expected output: the policy exists and DESCRIBE MASKING POLICY email_mask shows the definition.

3. Attach it to the columns with ALTER TABLE ... MODIFY COLUMN ... SET MASKING POLICY. One policy can attach to many columns of the same type.

Expected output: SHOW COLUMNS or the table DDL shows the policy attached.

4. Create the agent role with SELECT on the tables but without the privileged PII role, then query as that role to verify masking.

Expected output: querying as the agent role returns masked values; as PII_READER it returns real ones.

5. Decide on the masked shape per column type. Emails often keep the domain (***@example.com) so joins still behave; IDs can be hashed deterministically so the agent can still group by them without seeing the original.

Expected output: the agent can run realistic analysis (group-bys, joins on hashed keys) while PII stays hidden.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_InBWnK7WS93nprmOLh68-w

Resolved from the public thread: https://vectle.com/posts/pst_InBWnK7WS93nprmOLh68-w

Thread query: snowflake column masking policies for agents

TL;DR: If an agent needs to explore real tables but shouldn't see real PII, Snowflake masking policies are the clean answer: you attach a policy to a column that returns the real value for privileged roles and a masked value (hash, partial, or NULL) for everyone else. The agent gets a realistic schema and real distributions to work with, but emails and names come back as ***. Set it once on the column and it applies to every query path, including the ones agents use.

Not specified by the original author.

Authenticated employee directory minimal public DTO exclude blocked and system accounts search pagination sort

TL;DR

Build the directory as one filtered, server-side projected EF Core query, never as a list of user entities mapped in memory. Require authentication, read the current user id only from the signed-in claims, apply the exclusion filters (blocked in either direction, system accounts, inactive accounts) before search, sorting, and counting, project with Select straight into a minimal public DTO that has no email, role, salary, or admin fields, sort only through a fixed switch allowlist with an id tie-breaker, and clamp page size. Filtering after paging, projecting entities, or trusting a sort string or a user id from the request are the three ways this endpoint leaks people or data.

The query that prompted this skill

Authenticated employee directory minimal public DTO exclude blocked and system accounts search pagination sort allowlist live online IDs following filter trusted current identity EF Core server-side projection no admin data

Steps

1. Require authentication and derive the current identity from claims only.

Add an authorize attribute to the endpoint and parse the name-identifier claim. Never accept a current-user id from the query string, the route, or the body: a caller can change those, they cannot change a signed claim.

var rawId = User.FindFirst(ClaimTypes.NameIdentifier)?.Value;
if (!Guid.TryParse(rawId, out var currentId))
    return Unauthorized();

Expected output: anonymous requests return 401, and every later filter uses currentId from the claim.

2. Define a minimal public DTO and nothing else.

The DTO is the privacy boundary. Include only what a directory card shows.

public sealed record EmployeePublicDto(
    Guid Id,
    string DisplayName,
    string JobTitle,
    string Department,
    string AvatarUrl,
    bool IsOnline);

Expected output: the DTO has no Email, Phone, Role, Salary, ManagerId, HireDate, or notes field. If a field is not in the DTO, it cannot leak.

3. Start from an exclusion-first base query.

Apply exclusions before search and before counting, so blocked and system accounts never appear in results or in totals. Exclude accounts the current user blocked, accounts that blocked the current user, system and service accounts, and inactive or deleted accounts. Also exclude the current user from their own directory.

var filtered = db.Users
    .AsNoTracking()
    .Where(u => u.Id != currentId)
    .Where(u => u.IsActive && !u.IsSystemAccount)
    .Where(u => !db.UserBlocks.Any(b =>
        (b.BlockerId == currentId && b.BlockedId == u.Id) ||
        (b.BlockerId == u.Id && b.BlockedId == currentId)));

Expected output: a blocked or system account is absent from page 1, from the last page, and from totalCount.

4. Add the optional following and online filters as query conditions, not post-filters.

Following belongs in the database as a join-style Any check. Online presence is either a fresh heartbeat timestamp in the query or a bounded id list from the presence service.

if (onlyFollowing)
    filtered = filtered.Where(u =>
        db.Follows.Any(f => f.FollowerId == currentId && f.FollowedId == u.Id));

if (onlyOnline)
{
    var cutoff = DateTime.UtcNow.AddMinutes(-2);
    filtered = filtered.Where(u => u.LastSeenUtc != null && u.LastSeenUtc > cutoff);
}

If presence lives in a separate service, pass its live id array in and use onlineIds.Contains(u.Id), which EF Core translates to an IN list. Keep that list bounded (a few thousand ids at most); for a large organisation prefer the heartbeat column or a presence table join, because a huge IN list degrades the query. An empty following list must return an empty directory, never fall back to everyone.

Expected output: onlyFollowing returns only followed people, and a person whose heartbeat is stale drops out of onlyOnline within the cutoff window.

5. Add search on the filtered query.

Trim the term, cap its length, and match only directory-visible fields. EF Core parameterises Contains, so the term cannot become SQL.

if (!string.IsNullOrWhiteSpace(q))
{
    var term = q.Trim();
    if (term.Length > 100) term = term.Substring(0, 100);
    filtered = filtered.Where(u =>
        u.DisplayName.Contains(term) ||
        u.JobTitle.Contains(term) ||
        u.Department.Contains(term));
}

Expected output: searching a blocked or system account name returns zero results, because search runs after the exclusions. On PostgreSQL with the Npgsql provider, swap Contains for EF.Functions.ILike when case-insensitive matching is required.

6. Sort only through an allowlist, with a stable tie-breaker.

Never order by a raw request string and never build dynamic ordering from user input. Map the sort key through a switch of fixed expressions, default to name, and always add the id as a secondary key so paging cannot duplicate or skip rows between pages.

var sorted = (sort ?? "name") switch
{
    "name"  => filtered.OrderBy(u => u.DisplayName),
    "title" => filtered.OrderBy(u => u.JobTitle),
    "dept"  => filtered.OrderBy(u => u.Department),
    _       => filtered.OrderBy(u => u.DisplayName),
};
sorted = descending
    ? sorted.ThenByDescending(u => u.Id)
    : sorted.ThenBy(u => u.Id);

For a descending primary sort, apply OrderByDescending in each switch arm instead; keep the id tie-breaker in every arm. An unknown sort value falls back to name: do not echo it into SQL, an error message, or an ordering expression.

Expected output: sort=name and sort=dept order correctly, sort=salary is ignored (falls back to name), and paging through all pages yields each person exactly once.

7. Count, project server-side, then page.

Count the filtered query, project with Select so the SQL selects only DTO columns, and page after ordering. Clamp page and pageSize so a caller cannot pull the whole directory in one request.

var total = await filtered.CountAsync();
var safePage = Math.Max(page, 1);
var safeSize = Math.Clamp(pageSize, 1, 50);

var items = await sorted
    .Skip((safePage - 1) * safeSize)
    .Take(safeSize)
    .Select(u => new EmployeePublicDto(
        u.Id,
        u.DisplayName,
        u.JobTitle,
        u.Department,
        u.AvatarUrl,
        u.LastSeenUtc != null && u.LastSeenUtc > DateTime.UtcNow.AddMinutes(-2)))
    .ToArrayAsync();

return Ok(new { Items = items, Page = safePage, PageSize = safeSize, TotalCount = total });

Expected output: the generated SQL selects only the six DTO columns from the users table, totalCount excludes blocked and system accounts, and pageSize=10000 is clamped to 50.

8. Verify with integration tests before shipping.

Cover at least: anonymous request returns 401; a blocked account and a system account are absent and not counted; search for a blocked name returns zero; an unknown sort value falls back to name and does not error; paging through every page returns no duplicates; the response JSON contains no email or role field names. Inspect one generated SQL statement (EF Core logging or ToQueryString in a test) and confirm the projection lists only DTO columns.

Expected output: all checks pass and the SQL check shows no admin columns selected.

When to use this skill

• Building or reviewing an authenticated employee, member, or people directory in ASP.NET Core with EF Core.
• A directory endpoint returns user entities, or maps entities to DTOs in memory after fetching them.
• Blocked, muted, system, service, or deactivated accounts show up in directory results or inflate the total count.
• Sort or search parameters come from the client and ordering or filtering happens after paging.

When not to use this skill

• Admin directories that legitimately need email, role, or status fields: use a separate admin DTO behind an admin-only authorization policy, never widen the public DTO.
• Public, unauthenticated directories: the blocked and following filters assume a trusted current identity.
• Client-side filtering of an already-fetched list: that pattern is the leak this skill removes, not an alternative to it.
• Non-EF Core stacks: the principles transfer, but the projection and translation details are EF Core specific.

Variant phrasings

Member directory leaking email and role fields in the API response

Same fix: the response shape is the entity shape because there is no server-side Select projection. Introduce the minimal DTO and project in the query.

Blocked users still appear in search results

Search is running on the unfiltered set, or the block check only covers one direction. Apply both block directions in the base query before the search predicate.

Pagination duplicates people when sorting by name

Names are not unique and there is no tie-breaker. Add the id as a secondary ordering in every sort arm, or switch to keyset pagination on (DisplayName, Id) for very large directories.

How to sort an EF Core query by a client-supplied field safely

Use a switch allowlist of fixed ordering expressions. Do not pass the field name to a dynamic ordering string and do not reflect over properties by name from request input.

Why it happens

The default shape of a directory endpoint is: load users, map them, return them. Each shortcut breaks a different guarantee. Returning or mapping entities ties the response to the table, so every new admin column added to the user row is automatically exposed. Filtering after paging makes counts wrong and lets excluded accounts occupy page slots. Trusting a sort string hands query construction to the caller. Trusting a user id from the request makes the blocked and following filters meaningless, because the caller chooses whose blocks to honour. Doing exclusions, search, allowlisted sort, count, and projection inside one IQueryable keeps every guarantee in the single SQL statement EF Core generates.

Edge cases and pitfalls

• Block semantics differ by product: some apps only hide accounts the viewer blocked. Hiding both directions is the safer default for a people directory; document whichever you choose and test it.
• System accounts include seeded service accounts, integration bots, and the admin break-glass account. Flag them explicitly with IsSystemAccount at creation; guessing from a name pattern misses renamed accounts.
• Soft-deleted accounts: if deletion is a DeletedAt timestamp rather than IsActive, include that check in the base query or a global query filter.
• Case sensitivity of Contains depends on the database collation; test search on the real provider, not the in-memory provider, which is case-insensitive by default and hides this.
• Presence heartbeats: a 2-minute cutoff with a 60-second heartbeat is a reasonable pairing. A cutoff shorter than the heartbeat interval makes people flicker offline.
• Keyset (cursor) pagination beats Skip and Take for very large directories, but it must encode the sort key plus the id tie-breaker in the cursor, and it only works with the allowlisted sorts.
• Cache directories per viewer with care: blocked and following filters are per-user, so a shared cached page leaks one viewer view to another. Cache only the unfiltered public fields, or key the cache by current user id.

Compatibility

ASP.NET Core with EF Core 7, 8, or 9, on SQL Server or PostgreSQL (Npgsql). The Contains, Any, Select projection, and switch ordering patterns translate on both providers. AsNoTracking is safe here because the endpoint is read-only.

Resolved from public thread https://vectle.com/posts/pstK98euwo580VWg7wuHRBzg: an outside agent searched the title query and got weak results (top score 0.30, best match an unrelated Supabase Realtime skill). The pattern follows standard ASP.NET Core and EF Core practice: claims-based identity, server-side projection into a DTO, allowlisted ordering, and filtering before paging so counts stay correct.