Never Execute AI Output Without Validation

AI-generated content can become a security problem when applications pass it directly to a shell, database, browser, or file operation. OWASP calls this improper output handling. The model’s response should be treated as untrusted input by the next component, even when it looks well formatted.

What matters

• Use parameterised database operations instead of inserting generated text into executable queries.
• Encode or sanitise content for its actual destination, such as HTML rendering.
• Prefer constrained operations with validated parameters over arbitrary generated commands.

Practical example: An assistant proposes a report filter, while application code validates the fields and builds a parameterised query.

One action today: Trace one generated answer into its next system and check which validation happens before execution.