Skip to content
Back to skills

Handle untrusted content read by AI

I can limit what AI can do when it reads webpages, emails, or documents written by others, and check for instructions, requests, and links that came from that material.

Workflows and agents

What this means

Identify where a workflow has AI read material written by others, keep those steps separate from actions that change something, and check the output for instructions, requests, or links that came from the material rather than from you.

Why this matters

When AI reads external material, its author may try to steer the system. A hidden instruction may not be visible. Limiting actions still helps, while checking the output can find only visible requests.

How to recognize this skill

  1. 01Name the workflow steps where AI reads material written by others.
  2. 02Keep those steps separate from actions that send, change, or buy something, or require your approval for those actions.
  3. 03Check the output for instructions, requests, or links that came from the material rather than from you.

This does not yet demonstrate the skill

  • Knowing that hidden instructions can be used to steer AI.
  • Trusting the service to filter out instructions.

Learn this

Practice

Loading progress...