Skip to navigationSkip to loginSkip to content

This new phishing trick takes place out of your sight

Keytrade Bank logo

Keytrade Bank

keytradebank.be

November 13, 2026 

(updated November 16, 2026)

2 minutes to read

Phishing emails have always been written to be read by people. Now, however, more and more emails are addressing a second reader – an AI assistant that summarises or sorts your inbox, or even replies to emails. And fraudsters are hiding a separate message for the AI assistants that you can't see.

AI assistants are increasingly handling the tedious job of dealing with emails. Gemini is now built into Gmail, Copilot is integrated into Outlook, and ChatGPT or Claude can also be linked to your mailbox in just a few clicks. They summarise long threads, position important messages at the top and suggest replies. Some assistants can even send emails, schedule appointments or search for files.

Yet that convenience comes with a downside – anyone who lets an AI assistant read their emails also lets the assistant read emails sent by fraudsters. And fraudsters are no longer focusing all their attention on you.

One email, two targets

In early October, security firm Barracuda took a look at phishing emails that attempt to deceive humans and AI assistants alike. For human readers, the emails looked like ordinary internal correspondence, while for the AI assistants, the emails contained a second layer with hidden commands to manipulate the assistant into presenting the email as legitimate or urgent in its summary (source).

The idea behind it is as simple as it is smart. If you ignore the suspicious email, fraudsters will still try to get to you through the tool you do trust, as the AI assistant will list the phishing email as urgent.

Another example saw a hidden text block in an email containing an invoice. The text instructed the AI assistant to add a false 'priority action' in its summary, changing a supplier's payment details.

How exactly does it work?

An AI assistant works with commands, or more precisely 'prompts'. As an example, you can ask: "Summarise my emails today." The attacker uses a prompt injection to smuggle their own commands into the text the AI assistant has to process, hoping the assistant will execute the commands instead of simply reading them.

  • Scammers use a range of tricks to hide the prompts from you:
  • Text is hidden in the email's underlying code and never appears on your screen.
  • Text is made invisible, such as white on a white background or in a tiny font size.
  • Instructions are embedded in another component, such as image data.
  • Invisible characters are inserted between plain text.

You may well see what looks like an innocent email, but your AI assistant will be reading something completely different.

Why AI assistants are liable to falling for this trick

At the heart of the problem lies the fact that a language model doesn't make a hard distinction between text it needs to read and instructions it needs to follow. Everything is text. The UK National Cyber Security Centre (NCSC) warned in late 2025 that prompt injection may never be fixed and that the risks can only be limited (source).

EchoLeak, a flaw in Microsoft 365 Copilot, revealed in 2025 that this is not a theoretical danger. Researchers demonstrated that a single manipulated email was enough to manipulate Copilot into leaking sensitive information from an employee's mailbox, OneDrive or Teams, without the employee having to click on anything at all. Microsoft patched the flaw and saw no indications that it was actually being abused (source).

How big is the risk?

The size of the risk largely depends on what you let your AI assistant do. An AI assistant that only provides summaries can catch you off guard by flagging a phishing email as urgent or recommending a fake action. That sounds risky enough, but the decision remains your own.

An AI assistant that is allowed to act is another story entirely. If it can send emails, search through files or fill in forms, a hidden command can – in theory – let it do exactly what the fraudster wants, such as forward data, send a reply containing sensitive information or open a link. The more rights an AI assistant has, the greater the potential damage.

Something else is at play, too. People have learned not to trust suspicious emails, yet take very little time to reflect on what their AI assistant tells them. Recommendations from your own AI assistant feel neutral and reliable – and it is that trust and confidence scammers try to exploit.

Meanwhile, suppliers of AI tools continue to build in security features such as filters that can recognise and remove hidden content. However, it is still a game of cat and mouse – and you therefore form part of the solution for staying protected.

How can you protect yourself?

1. A summary is not confirmation

If your AI assistant advises you to pay an invoice, change details, log in or share something, always open the original email in the first instance and check who sent it. Never rely purely on the summary.

2. Be wary of urgent flags from your AI assistant

Has your AI assistant suddenly marked an email from an unknown sender as urgent or suggested an action you didn't ask for? This is just one reason to be extremely vigilant, particularly when it comes to money or personal data.

3. Give your AI assistant as few permissions as possible

An AI assistant that summarises your emails doesn't need to be able to send them on its own. Check the settings to see which access rights you have enabled, and disable anything you don't need. Avoid linking your AI assistant to all your services at once – the fewer services it can reach, the less it can leak.

4. Maintain control over important actions

Never let your AI assistant send emails, prepare payments or share data automatically. Where possible, select an institution where you first have to approve each action – and read what you approve carefully.

5. Check changes to payments using a separate channel

If an email or your AI assistant informs you that your energy supplier's account number has changed, call the supplier yourself using a number you already have. Never use the contact details listed in the message itself.

6. Only use official applications

Only link your mailbox to AI tools and integrations from well-known suppliers. An unknown extension or app that promises to make your inbox 'smarter' will gain access to all your emails, including emails from your bank.

7. Never let your AI assistant handle your banking

Never give an AI assistant your passwords, security codes or access to your banking environment. Manage your banking matters yourself using your bank's official app or website.

8. If you notice anything strange, stop what you're doing immediately

If your AI assistant refers to instructions you never issued, or it suddenly wants to do something that has nothing to do with your question, stop what you're doing, disconnect the AI assistant from your mailbox if necessary and forward the suspicious email to verdacht@safeonweb.be.

Still had something go wrong?

If you transferred money or data after receiving advice from your AI assistant, contact your bank immediately. At Keytrade Bank, you can call +32 (0)2 679 90 00. You can also block your payment cards easily in the Keytrade Bank app or by calling Card Stop on +32 (0)78 170 170. What you should do next can be found in our step-by-step guide for what to do if you've fallen victim to fraud. You can find out more about how fraudsters are using AI in our article on AI fraud.

Stay safe online with Keytrade Bank

At Keytrade Bank, we make security a top priority. If your bank details have been listed on a suspicious website, your personal details have been shared with an unknown person over the phone or you've spotted an unknown payment that you didn’t make, call us immediately on +32 (0)2 679 90 00.