Initech's email assistant reads the inbox and can send email on the user's behalf. Your call: which defenses do you ship, knowing each one costs something?
Key idea: Prompt injection hides instructions inside content the model reads. Check what comes back from tools, not just what the user types, and limit what the agent is allowed to do.
Things to try
- Ask for a summary with the defaults. The input guardrail is on, but the attack is hiding in an email, not in your message.
- Turn on Scan tool results and ask again. The poisoned email never reaches the model.
- Scanning off, Mark untrusted content on. Ask for a summary, then for action items. The model is told that emails are data, not orders.
- Mark untrusted content off, and set Send email to Ask first. Even a fooled agent can't send anything without you.
- Set Send email back to Allowed and send the last prompt, a direct attack. This is the case the input guardrail is built for.
Try a prompt