How do I protect my AI from prompt injection?

Treat everything the AI reads that you did not write, such as web pages, documents, emails and tool descriptions, as untrusted. Limit what the AI is allowed to do on its own, check the tools it can reach, require approval for high-impact actions like payments or deletions, and keep a record of every action so anything that slips through is traceable.

How it works

1. Check the tools your AI connects to for hidden instructions (MCP Shield does this).
2. Give agents the least privilege they need, and gate payments behind a signed approval.
3. Validate inputs against their claimed source before trusting them.
4. Seal each action so you can prove afterwards exactly what happened.