> For the complete documentation index, see [llms.txt](https://hacking-notes.jord4n.pro/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://hacking-notes.jord4n.pro/web/llm-web-attacks/exploiting-unsafe-llm-output-handling.md).
# Exploiting Unsafe LLM Output Handling
Exploiting Unsafe LLM Output Handling notes for PortSwigger Web Security Academy labs, including exploitation workflow, payload testing, Burp Suite validation, and remediation context.
### Exploiting Insecure Output Handling in LLMs
#### Context
This lab displays the output of the LLM in an unsecured manner, allowing a **XSS**. User **carlos** often uses live chat to request product information **Lightweight "l33t" Leather Jacket**.
#### Objective
Perform an indirect \*\* injection (indirect quick injection)\*\* to execute a **XSS** which **removes the carlos account**.
### 1) Setup
* Create a user account.
* Access to **live chat** with LLM.
### 2) XSS Verifications
#### Cat Test (vulnerable)
Test Payload (set an alert):
```html
```
#### Test on Comments (not Directly Vulnerable)
* Comments do not appear to perform HTML/JS in the same way.
We therefore go through a **source consulted by the LLM** (recovered contents and displayed in the chat).
### 3) Point of Entry: Content of Opinions Consulted via LLM
Ask the LLM to display product reviews, for example:
```bash
View reviews of the ZZZZZZ bed: your new home office.
```
Observed response (example):
```bash
4. **User** - 25 January 2026
- *Review:* *(Review contains an image)*
```
The idea is to inject HTML into a notice so that it is then **replaced in the live chat** (where the display is vulnerable).
### (4) Targeted Action: Account Deletion via Form
To delete an account from My account, trigger:
```bash
document.forms[1].submit()
```
Formulaire correspondant :
{% code overflow="wrap" %}
```html
