Every developer knows the rule: never pass user input to innerHTML. Yet as LLMs move into our applications, we hand model output straight to the DOM, the shell, or the database, because it came from our own system and therefore feels trustworthy. It isn't.

This talk looks at OWASP LLM05, improper output handling, and what it means for web security. We'll walk through cases where entirely innocuous-looking input steers a model into producing XSS and injection payloads, and see why the input filters you already have never fire.

You'll leave professionally pessimistic about model output: how to sanitize it, how to render generated Markdown safely, and how to treat AI content as the untrusted input it always was.