Hidden instructions: when a document tries to boss your AI around
Emails, documents and web pages can contain instructions aimed at the AI reading them — sometimes invisible to you. It’s called prompt injection.
- Be careful asking AI to summarise and act on content from outside your business — emails, attachments, websites.
- Don’t let an AI tool send, reply, share or buy anything based on outside content without you checking first.
- If an AI summary suddenly tells you to click a link, reset a password or contact someone, treat it like a phishing email.
- Watch for answers that go off topic, push you to act urgently, or ignore what you asked.
- Report odd AI behaviour to IT, with the document or email that caused it.
Based on: OWASP — LLM01:2025 Prompt injection · ASD’s ACSC — Engaging with artificial intelligence
Still stuck after those?
You’ve done the right thing by trying. Send whedo.it a quick note — what you tried, what you saw — and a senior pair of eyes will be on it the same business day.