Skip to main content

Article

AI Email Assistant Phishing: How Hidden Text Attacks Work

AI email assistant phishing hides instructions your AI reads but you can't see. How the attack works and five rules to protect a small team now.

Andy Robbins5 min read
  • ai
  • security
  • email
  • small business

AI email assistant phishing is a new twist on an old scam. A phishing email still tries to fool you, but now it also carries hidden instructions meant for the AI tool that reads your inbox. You never see those instructions. Your assistant does.

If your team uses AI to summarize email, draft replies, or sort the inbox, this one matters. Security researchers at Barracuda described the attack this month, and the defenses are simple enough that a small business can put them in place this week.

How AI email assistant phishing works

Barracuda published two posts on this in October 2026. The first describes emails built to target a person and an AI assistant at the same time. The second goes deeper on the technique, called indirect prompt injection.

Malwarebytes summarized one example well. For the human, the email had a password-protected attachment with the password in the message. For the AI assistant, it had hidden instructions meant to change how the assistant summarized or handled the email.

Malwarebytes put the risk plainly: you ignore a suspicious email, then ask your AI assistant to summarize your inbox, and the message you avoided comes back as an urgent recommendation from a tool you trust.

Where the hidden instructions hide

According to Barracuda's research, attackers hide the instructions using:

  • HTML comments. Text in the email's code that doesn't display on screen.
  • Invisible text. Styling that makes words white on white, tiny, or hidden.
  • Base64. Text encoded so it looks like random characters to a person.
  • Zero-width characters. Characters that take up no space and can't be seen.

The goal, per Barracuda, can be getting the assistant to mark a message as urgent, change its recommendations, leak data, or trigger an action.

Why small businesses are exposed

Big companies have security teams filtering email before it reaches anyone. Many small businesses don't. And small teams are often the fastest to adopt AI assistants, because they save so much time.

The risk grows with permissions. An assistant that only summarizes email can mislead you. An assistant that can send email, pay invoices, or change vendor details can do real damage on its own.

This is the same reason we recommend a simple AI policy for small businesses before tools get connected to everything.

5 rules to protect your team

Barracuda's guidance boils down to treating outside email as untrusted and keeping a human in the loop. Here's how to apply that in a small business:

1. Limit what your AI assistant can do. Give it the least access it needs. If it only needs to read and summarize, don't let it send, delete, or forward.

2. Require human approval for anything with money. Payments, refunds, invoices, and changes to bank or vendor details should always need a person to approve, no matter what the assistant recommends.

3. Verify through a separate channel. If an email asks for a payment change or urgent action, call the person using a number you already have. Don't reply to the email, and don't use contact details from it.

4. Be skeptical of "urgent" in a summary. If your assistant flags something as urgent that came from an unknown sender, open the original email and read it yourself.

5. Review connected tools regularly. Every few months, check which apps and browser extensions have access to your email. Remove what you don't use. Our post on the ChatGPT Chrome extension covers what browser-based AI can see.

Train the team in 15 minutes

You don't need a long training. Share three points at your next team meeting:

  • AI summaries can be manipulated by the emails they summarize.
  • Anything involving money or account changes gets verified by phone.
  • If a summary feels off, read the original.

If you're rolling out AI across the team, our guide on how to get your team to use AI covers how to build habits that stick, including safe ones.

A quick checklist for this week

If you only have half an hour, do these:

  1. List every AI tool that touches email. Include inbox assistants, browser extensions, and any automation that reads incoming messages.
  2. Write down what each one can do. Read only, draft, send, or take actions in other apps.
  3. Turn off anything you don't need. If nobody uses the auto-send feature, switch it off.
  4. Add an approval step for payments. In your accounting or banking tool, require a second person to approve new payees and bank detail changes.
  5. Tell your bookkeeper or office manager. They're usually the person a fake "urgent invoice" is aimed at.

None of this needs new software. It's mostly settings and a short conversation.

Should you stop using AI email assistants?

No. They save real time, and the fix isn't to avoid them. It's to give them the right limits. An assistant that reads and drafts, with a person approving anything that leaves the building or moves money, keeps almost all the benefit with far less risk.

How we think about AI tools at AI Answered

We set up AI tools for established companies as part of a bigger growth system, and approval steps are built in from the start. At AI Answered, we connect AI to sales follow up, paid social, and tracking, and we keep a human check on anything that sends, spends, or changes data. That's slower by a few seconds and safer by a lot.

Common questions

What is AI email assistant phishing? It's a phishing email that includes hidden instructions meant for an AI assistant, trying to change how it summarizes or acts on the message.

Can I see the hidden instructions? Usually not. They're hidden with tricks like invisible text, HTML comments, or zero-width characters.

Which AI email assistants are affected? Any assistant that reads email content can be targeted. The protection is the same: limit permissions and require human approval for important actions.

What's the single most important defense? Never let an AI assistant move money or change payment details without a person approving it.

Ready to put this into your business? We map how you run, build what you approve, train your team, and prove the hours and dollars saved.

All articles

Want this built into your business?

Three questions, 20 seconds. If we can help, we start with a review of your numbers.