Advisory

GHSA-pv5w-4p9q-p3v2

Details, part 2 of 2. The advisory

Details

The developer's intent: only top-level keys (nick, tagline) are ever requested. internal is private bookkeeping.

$ node sqlite-runtime.mjs
===== Legitimate request =====
userInput = "nick"
  compiled SQL:  select "profile"->>'$.nick' as "value" from "me" where "id" = ?
  result:        [ { value: 'alice' } ]

===== Injection: dot lets attacker reach nested "internal" object =====
userInput = "internal.ssn"
  compiled SQL:  select "profile"->>'$.internal.ssn' as "value" from "me" where "id" = ?
  result:        [ { value: '111-11-1111' } ]

userInput = "internal.token"
  compiled SQL:  select "profile"->>'$.internal.token' as "value" from "me" where "id" = ?
  result:        [ { value: 'tok_abcdef' } ]

userInput = "internal.admin"
  compiled SQL:  select "profile"->>'$.internal.admin' as "value" from "me" where "id" = ?
  result:        [ { value: 1 } ]

Expected vs. actual: the application invariant was "the user can only read top-level keys of their profile". The output violates that invariant — internal.ssn, internal.token, and internal.admin are returned even though internal was never meant to be addressable through this endpoint.

The same pattern is exploitable on MySQL (where * and ** wildcards make it strictly worse — a single * enumerates every sibling at the current level in one row) and on PostgreSQL when using the ->$/->>$ operators (which target MySQL-style JSON-path strings on PG ≥ 17 / via jsonb_path_query).

Impact

  • Authorization bypass on JSON sub-fields. Any kysely-built query whose JSON-path key/index argument is partially or fully attacker-controlled — even in fully type-safe code where the column type is Record<string, T> — leaks data the developer believed was scoped behind the explicitly-listed key. SSNs, tokens, admin flags, internal IDs, anything stored as a nested member of the same JSON document is reachable.
  • Wildcard reads on MySQL / PostgreSQL `->$`. key('*') compiles to '$.*', returning the array of every value at the current depth in one round-trip. key('**') recurses across the whole document. The fix does not strip either token.
  • Write access in update statements. Kysely uses the same path compiler for update().set(eb => eb.ref(col, '->$').key(input), value)-style writes (and jsonb_set helpers). An attacker who can drive both the path and the value can therefore write into nested fields they should not be able to set — for example flipping an admin flag or rewriting a nested role.
  • Bypasses the recently-fixed precedent. The maintainers shipped commit 0a602bf (PR #1727) specifically to harden this surface. That fix removed the ' (quote) primitive but left every JSON-path metacharacter alone, so the surface is still open against any caller that *thought* it was now safe.
  • Practical bounding. The attacker needs a code path where a request-derived string lands in .key(...) or .at(...). This is a recognised pattern (filter-by-field, dynamic select for admin dashboards, Strapi-style JSON-blob columns); it is not a default kysely behaviour but is plausibly common. The vulnerable path is also exercised any time a developer writes db as Kysely<any> (covered by the older GHSA-wmrf-hv6w-mr66 advisory) — but unlike that advisory, the bug here triggers in fully-typed code on Record<string, T> columns.

Suggested fix

Treat path legs as a structured emission, not a string-literal escape. The narrowest safe patch is a dedicated sanitizeJSONPathLeg that only emits a known-good character set per leg type and rejects everything else, since JSON-path quoting differs by dialect (MySQL allows "…"-quoted member names; SQLite is more permissive but still has a grammar; PostgreSQL jsonpath is strict).

// src/query-compiler/default-query-compiler.ts
const JSON_PATH_MEMBER_OK = /^[A-Za-z_$][A-Za-z0-9_$]*$/

protected override visitJSONPathLeg(node: JSONPathLegNode): void {
  if (node.type === 'ArrayLocation') {
    this.append('[')
    if (typeof node.value === 'number') {
      this.append(String(node.value | 0))      // int-coerce
    } else if (node.value === 'last' || /^#-\d+$/.test(node.value)) {
      this.append(node.value)                  // documented dialect tokens
    } else {
      throw new Error(`invalid JSON array index: ${node.value}`)
    }
    this.append(']')
    return
  }
  // Member
  this.append('.')
  if (typeof node.value !== 'string' || !JSON_PATH_MEMBER_OK.test(node.value)) {
    // Per-dialect quoted-member escape would go here; default = reject.
    throw new Error(`invalid JSON path member: ${JSON.stringify(node.value)}`)
  }
  this.append(node.value)
}

For dialect-specific behaviour (MySQL "…"-quoted members, SQLite bracket-quoted), each dialect compiler should override the helper and apply the appropriate quoting + double-the-quote rule, the same way sanitizeIdentifier already does.

Consider also: parameterise JSON paths whenever the dialect supports it (PostgreSQL jsonb_path_query($1, $2), MySQL JSON_EXTRACT(?, ?)), so attacker-controlled keys are bound, not concatenated. Add a regression test to test/node/src/json-traversal.test.ts asserting that eb.ref('c','->$').key('a.b').compile().sql is either rejected, or emits MySQL '$."a.b"' / SQLite '$.["a.b"]' (quoted-member form), and explicitly differs from key('a').key('b').

A backstop hardening: tighten the .at() runtime to accept only number | 'last' | '#-${digits}' (matching the type signature), and tighten .key() to only accept strings that match keyof O at runtime when O is statically known.

Sources & licences