Insecure output handling

Insecure output handling happens when software takes whatever an AI model writes and passes it to another system without checking or cleaning it first. That other system might be a web page, a database, a command on a computer or an email. An attacker can often steer what the model writes, for example through prompt injection, so the unchecked answer can carry an attack.

Here is an example. A chatbot on your donation page is tricked into replying with a hidden script. If the page shows that reply without checking it, the script runs for every visitor who sees it.

The OWASP Top 10 for LLM Applications 2025 lists this risk as Improper Output Handling (LLM05:2025). Our plain-language guide to that list covers the other risks.

Why it matters

Some AI tools do more than chat. They post to your website, update your donor records or send emails for you. When a tool works like that, treat its answers like text from a stranger. Before you connect it, ask the vendor: “Do you check and clean the AI’s output before it reaches our website, database or inbox?” If the vendor gives a vague answer, slow down. The risk grows when a tool also has excessive agency.