How do you build an AI feature that reads text from strangers without letting that text take control?
- Module
- 03/05
- To read
- ~9 min
A support assistant could read customer tickets and send emails. One ticket contained hidden instructions. What do you think happened?
In the module you make your guess first, then read what happened and why.
Treat All Model Input as Untrusted
Least Privilege for Every Tool
Approval Before Consequential Actions
Validate Output and Limit Where Data Can Go
A risk map for your AI feature
A spec for your own app, as a text file you download. The module also gives you the prompt for Claude Code, Codex or Antigravity, a checklist to check what it built, and recommended skills.
- Authentication and authorization
- Rate limiting
- Observability