Improper Output Handling
Model output is trusted and passed unvalidated into something that executes it — rendered as HTML (XSS), run as SQL, or sent to a downstream system.
What it means
Improper output handling is treating the model’s response as trusted and feeding it, unchecked, into something that acts on it: rendering it as HTML in a browser, running it as a database query or shell command, or forwarding it to another API.
The model’s output is just text — text an attacker may have influenced through injection. Executing it blindly is what turns a prompt injection into real code execution.
Why it matters
It fell to #10 in the ranking but remains a critical link: it is the step that converts a manipulated model into a concrete breach in the surrounding system.
How we approach it
The assistant’s output is rendered as text, never as executable markup — responses are placed into the page as safe text nodes, so neither the model nor an injection can emit script that runs in your users’ browsers.
Links are validated to safe schemes before they are made clickable; anything else is shown as plain text.
Actions the model proposes go through your declared, typed endpoints — not arbitrary strings handed to an interpreter — and changes are confirmed before they run.
Shared note: the application embedding the assistant should preserve the same discipline on its own side — treat model output as untrusted wherever it flows next.
Frequently asked
Can the assistant’s output run code in my users’ browsers?
No. Output is inserted as text, not as live HTML, so a response cannot inject a running script. Links are also restricted to safe schemes before becoming clickable.
What about actions the assistant triggers downstream?
They flow through your declared, typed endpoints rather than arbitrary commands, and changes require confirmation — so output is never handed straight to an interpreter.