Use AI to fight AI

Scammers use AI to write cleaner, more convincing messages. You can use the same kind of tool to slow down, look closer and make a better call. The point is to have a second pair of eyes, one that never gets tired and never feels rushed, while you keep the final decision.

Used well, an AI assistant helps good people work together: you, your colleague who forwarded the odd invoice, the IT team who will block the sender. Used carelessly, it can leak the very data a scammer wants. The difference comes down to a few habits.

Why the old tells are fading

For years, the advice was to look for bad spelling and clumsy grammar. That advice is aging fast. The U.S. Cybersecurity and Infrastructure Security Agency (CISA) notes that AI can now produce messages with flawless grammar, so spelling mistakes are no longer a reliable sign. The FBI warned in December 2024 that criminals use generative AI to make fraud more believable and easier to run at scale.

What still works is looking at what a message is trying to make you do. The UK National Cyber Security Centre (NCSC) lists the classic pressure points: a claim of authority, urgency or a deadline, a strong emotional pull, something scarce on offer and a hook tied to current events or the time of year. An AI assistant is good at spotting those patterns quickly, especially when you are busy.

Three safe ways to ask for help

Before you paste anything, strip out personal and company details (more on that below). Then try one of these.

Ask for the red flags

Paste the text of the message, with names and account details removed, and ask something like:

List the warning signs in this message. Point to the exact words that create urgency, claim authority or ask me to act outside normal channels.

A good answer names specific phrases. Use it to sharpen your own judgment rather than as a verdict.

Ask it to explain a link without clicking it

Copy the link address as text (right-click and copy, or long-press on a phone) and ask the assistant to break it down: what the real domain is, whether the brand name appears only as a subdomain or in the path, whether the spelling is slightly off. Never open the link to "see what happens." CISA's advice is to reach the company by typing its address yourself or searching for it.

Ask it to draft a call-back question

This is the most useful one. If a message claims to come from your bank, your CEO or a supplier, ask the assistant to help you write a short, neutral question you can use when you contact them through a number or address you already trust. For example: "I got a request to change your bank details. Can you confirm that came from you?" The NCSC gives the same advice: skip any contact details printed in the message and use the ones on the organization's official website.

What never goes into a chatbot

The NCSC advises against putting sensitive information into queries to public AI tools, or anything that would cause problems if it became public. Queries are visible to the company that runs the service, may be stored and may be used to develop it later. Read the provider's terms before you share anything sensitive.

In practice, keep these out of any assistant your organization has not approved for them:

• Passwords, one-time codes, recovery phrases and API keys

• Customer data: names, addresses, account numbers, health or payment details

• Confidential documents such as contracts, financials, board papers and source code

• Personal details about colleagues or family members

You can still get help with a suspicious message. Replace real names with "Person A," swap the account number for "XXXX" and remove signatures. The red flags survive the redaction.

A word of caution: the message may talk back

One more twist is worth knowing. A scam email can contain hidden text aimed at the AI, not at you, telling it to call the message safe. This is called prompt injection, and the OWASP Top 10 for LLM Applications ranks it as the top risk for AI applications in its 2025 list, including instructions hidden in outside content that humans can't see.

So treat the assistant's answer as advice. If it says "looks fine" but your gut says something is off, your gut wins, and you make the call-back anyway. AI can also simply be wrong. The NCSC's guidance for leaders lists inaccurate or made-up answers among the known risks of these tools.

Team rules worth writing down

In many organizations, people already use AI at work, with or without a policy. A short written agreement keeps everyone safe and takes the guesswork out. The NCSC makes an important point here: the burden of using AI safely should not fall mainly on individual users. Leaders set the tools and the rules so people can do the right thing easily.

A one-page policy can cover:

1. Approved tools. Which AI assistants people may use for work, and which accounts (company accounts only).

2. The never-paste list. The categories above, in plain words, with examples from your own business.

3. The suspicious-message routine. Redact, ask the assistant for red flags, then verify through a known channel before acting.

4. Where to report. One inbox or chat channel for anything odd, with a promise of a thank-you rather than a telling-off. In the UK, the NCSC also accepts suspicious emails at report@phishing.gov.uk.

5. A pause rule for money and data. Any request to pay, change bank details or share data gets a second person and a call-back, no matter how real it looks.

At home, the same idea works in a simpler form. The FBI suggests families agree on a secret word or phrase to confirm who is really calling, and hang up and call back on a number you look up yourself. Show your parents or teenagers how to paste a strange text into an assistant and ask, "What here is trying to rush me?" It turns a scary moment into a shared habit.

The real advantage is people

AI won't stop every scam, and it doesn't need to. What helps most is a culture where anyone can say "this looks odd, can you check it with me?" and get a quick, kind answer. An assistant is one more helper in that conversation. The judgment, the call-back and the decision stay with you and your team.

Sources

• U.S. Cybersecurity and Infrastructure Security Agency, "Recognize and Report Phishing," undated (archived page): cisa.gov

• FBI Internet Crime Complaint Center, "Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud," Dec. 3, 2024: ic3.gov

• UK National Cyber Security Centre, "Phishing scams: how to spot and report them," reviewed Sept. 5, 2022: ncsc.gov.uk

• UK National Cyber Security Centre, "Phishing scams: how to spot and report them (report a scam email)," accessed Oct. 25, 2026: ncsc.gov.uk

• UK National Cyber Security Centre, "ChatGPT and large language models: what's the risk?," March 14, 2023: ncsc.gov.uk

• UK National Cyber Security Centre, "AI and cyber security: what you need to know," Feb. 13, 2024: ncsc.gov.uk

• OWASP Gen AI Security Project, "LLM01:2025 Prompt Injection," 2025: genai.owasp.org

Or let NimBot handle it

You can do all of this yourself, and it works. If you'd rather not run it alone, NimBot handles the training, the drills and the reminders for your whole team. Try it at app.nim.bot.