The v2.3.0 fix (commit e88dabb) routed the _select field through a strict new gate, sanitizeSelectField (postgres.go:947/1003/2097). The sibling groupby path was left on the far weaker isSafeSQLExpression. adapters/postgres/postgres.go, GroupByClause (line 1946): a groupby field that contains balanced parens takes the raw-SQL-expression branch (around lines 1999-2006), gated only by isSafeSQLExpression (line 2001, defined at 2043). That gate validates only the OUTERMOST function name against a 13-entry allowlist and uses a character allowlist that permits spaces, single quotes and parens. It blocks ;, --, /* , unbalanced parens and an outer pg prefix, but NOT a trailing UNION SELECT ..., a nested (SELECT ...) subquery, or an inner pg function. sanitizeSelectField (the CVE fix) rejects exactly those.
GET /{database}/{schema}/{table}?_select=id&_groupby=upper((SELECT+1))+UNION+SELECT+current_user. Sink trace: controllers/crud.go:131 calls GroupByClause unconditionally, concatenates it into sqlSelect (crud.go:133), executed at crud.go:176 with only the WHERE values parameterized. Reached after only a table read-permission check, passable unauthenticated in the default config (the same pipeline and attacker model as the anchor).
EXECUTED with negative controls (agent-run go test ./adapters/postgres/, PASS): isSafeSQLExpression wrongly accepts four payloads and GroupByClause emits them raw (e.g. GROUP BY upper((SELECT 1)) UNION SELECT current_user, and GROUP BY abs((SELECT 1 FROM pg_sleep(3)))); negative controls with ; and -- are rejected; and sanitizeSelectField (the CVE fix) rejects the exact UNION payload the sibling gate accepts. I independently re-confirmed at HEAD that _groupby is gated by isSafeSQLExpression (postgres.go:2001) while _select uses sanitizeSelectField (postgres.go:947). NOT verified: end-to-end UNION exfiltration against a live Postgres (needs column/type alignment, Postgres-gated). The pg_sleep DoS primitive is unconditional and needs no alignment; char-by-char blind extraction is constrained (commas are pre-split, and =/</> are blocked).
Route _groupby through the same strict gate as select (sanitizeSelectField-style: reject nested subqueries, UNION, inner pg functions, and anything past a single allowlisted function call), or parameterize/whitelist the group-by columns.
The other builders (WhereByRequest, OrderByRequest, JoinByRequest, CountByRequest, the jsonb/tsquery paths, ->>having, ident.IsValid) are sound; only the _groupby non-having function branch is the weak sibling.
Five-source enumeration: CVE-2025-58450 (GHSA-p46v-f2x8-qp98, "Systemic SQL Injection", GO-2025-3941) and the newer GHSA-qvx3-q8vx-9q3c (which covers _select, fixed v2.3.0). Neither mentions _groupby / GROUP BY. The systemic advisory's own "nested queries can be performed" note supports the incomplete-fix framing. Not contested.
Dedup gate clean: no prest/prest blocking verdict in COVERED-INDEX, DISCLOSURE-INDEX or SCOUT-EXCLUDE.
AI assistance wsa used while investigating. The injection is executed-verified with negative controls (isSafeSQLExpression accepts four UNION/subquery/pg_sleep payloads that GroupByClause emits raw, while ;/-- are rejected and the _select gate rejects the same UNION payload), and I confirmed at HEAD that _groupby uses the weak isSafeSQLExpression while _select was hardened with sanitizeSelectField.
{
"cwe_ids": [
"CWE-89"
],
"github_reviewed": true,
"github_reviewed_at": "2026-11-20T14:41:42Z",
"nvd_published_at": "2026-11-19T19:17:17Z",
"severity": "HIGH"
}