Skip to content

T-018 · Permissions

Don’t let AI take orders from what it reads.

Do this

Know that emails, web pages and documents can hide instructions aimed at AI tools.
If an AI tool reads outside content, don’t let it also send, pay or share anything without your OK.
Instead of asking an assistant to “handle” a supplier’s email, ask it to summarize the email and draft a reply for you to review.

Next tip: T-019 →

Sources

NIST Generative AI Profile (NIST AI 600-1), section 2.9, Information Security: indirect prompt injection planted in data an AI retrieves, used by researchers to steal proprietary data
IMDA Singapore, Model AI Governance Framework for Agentic AI, Version 1.5 (2026), section 1.2.1, prompt injections can manipulate an agent into leaking the data it has access to (p. 10)