High

XSS Payload Injection in Tool Response Output

An AI agent can call external tools or skills and pass their output back to a chat window or dashboard. If that output contains script tags, event handlers, or javascript: links and the interface renders it without sanitizing it, the code executes in the user's browser just like a classic cross-site scripting attack.

How the attack works

An attacker controls or poisons a tool, plugin, or data source that an AI agent queries. The tool's response contains an XSS payload — a script tag, an onerror handler, or a javascript: URI — instead of plain data. The agent forwards that response to a web-based chat UI or dashboard as-is. Because the frontend renders the content as HTML rather than escaping it, the injected code runs in the victim's browser session.

Why it matters

A working XSS in an agent's UI can steal session tokens, hijack the user's account, or take actions on their behalf inside the agent dashboard, using a trusted-looking tool response as the delivery mechanism.

What you can do

  • Treat all tool and skill output as untrusted input and HTML-escape or sanitize it before rendering in any web UI.
  • Use a strict content security policy that blocks inline scripts and javascript: URIs in the agent interface.
  • Render tool output as plain text or through a sanitizing markdown renderer rather than raw HTML.
  • Review any tool, plugin, or third-party data source that can inject content directly into the agent's displayed responses.

Known benign look-alikes

  • Security audit reports listing XSS examples as findings (without live HTML rendering)
  • Web development documentation explaining XSS in code blocks
  • CTF challenge writeups containing harmless alert() demos in non-rendering contexts

References