CodeQL 2.26.0: What Does the System Prompt Injection Query Catch in AI Applications?

Short answer

System Prompt Injection occurs when untrusted user inputs alter or override the system instructions guiding an AI model's behavior, leading to unauthorized tool execution or sensitive data disclosure.

Static analysis tools like **CodeQL** incorporate security rules (e.g., CodeQL 2260 query patterns) to detect un-sanitized user inputs concatenated directly into system prompt instructions.

Defensive Pattern:

// VULNERABLE: Direct string concatenation into system prompt
const prompt = `System: You are an assistant. User query: ${userInput}`;

// SECURE: Separating system instructions from user messages
const messages = [
  { role: 'system', content: 'You are an assistant. Do not execute system commands.' },
  { role: 'user', content: userInput }
];

Prompt Injection Checklist

• [ ] Separate system instructions from user input in message arrays
• [ ] Run CodeQL static analysis to detect prompt injection vectors
• [ ] Validate and sanitize all external user inputs before LLM invocation