Hidden instructions: why AI connected to your email needs care

AI tools are increasingly connected to your email, calendar, files and browser, so they can read and act for you. That's powerful, but it brings a new risk. A web page, email or document can contain hidden text written to steer the AI, for example telling it to send information somewhere or to ignore your instructions. This is called prompt injection.

How to use connected AI safely

• Only connect approved tools to work email, calendar or files, and ask us before connecting anything new.
• Give it only the access it needs. If a tool just needs to read your calendar, don't let it send email.
• Read what it's about to do before you approve an action such as sending an email, sharing a file or filling in a form.
• Be wary of surprises. If the AI suddenly suggests sending data somewhere, visiting a site or changing settings you didn't ask for, stop.

Point it at trusted content

Ask AI to summarize or act on documents and sites you trust. Be careful letting it browse unknown websites or process emails from people you don't know on your behalf.

Something unexpected happened? Stop the task, disconnect the tool if you can, and call us on 707-755-8000.