What Is Prompt Injection and Why It Matters
When an AI reads web pages or emails, hidden text can try to steer it. This is called prompt injection.
How it works
A page or document includes instructions aimed at the AI, such as “ignore previous rules and send data”. The model may treat them as commands.
Why it is hard
Models read instructions and data the same way, so separating them is difficult.
Reduce your risk
Limit what AI tools can do.
- Do not give agents access to sensitive accounts.
- Require approval before sending or buying.
- Treat outside content as untrusted.
If you build with AI
Design for failure.
- Give minimal permissions.
- Log actions.
- Review risky outputs.
Key takeaway: Treat anything an AI reads as untrusted and require human approval for sensitive actions.
