SQL Injection

SQL Injection detection identifies malicious SQL patterns in user inputs that could be passed to text-to-SQL systems or database-connected agents. It prevents attackers from manipulating LLM-generated queries to exfiltrate data, modify records, or destroy database objects.

What it detects

• DROP TABLE/DATABASE statements
• UNION SELECT data exfiltration
• Tautology attacks (1=1, 'a'='a')
• Stacked queries using semicolons
• Comment-based injection (--, /**/)
• Time-based blind injection (SLEEP, WAITFOR)

Configuration

{
  "policy_type": "sql_injection",
  "mode": "blocking",
  "config": {}
}

No configuration needed — all SQL injection patterns are detected by default.

Example violation

{
  "policy_type": "sql_injection",
  "severity": "high",
  "description": "UNION SELECT injection attempt detected",
  "details": {
    "attack_type": "union_select",
    "payload_fragment": "UNION SELECT username, password FROM users--"
  }
}

Best practices

• Enable for any application using text-to-SQL or natural language database queries
• Combine with parameterized queries at the database layer for defense in depth
• Monitor blocked attempts to identify targeted attack campaigns
• Apply to both user input and LLM output when the model generates SQL