How to spot AI scams before they cost you money

On October 29, thieves took about $6.3 billion from Binance. The exchange's own report says the attack began months earlier, with a tampered download of an AI coding tool and a fake job interview. The same day, a hacked website script cost Polymarket users about $2.7 million. Nobody has shown that AI ran either attack, but they landed in a year when Microsoft, Google and the FBI all documented criminals using it. So how to spot AI scams is now a practical question for anyone who sends invoices, pays suppliers or opens files from strangers.

You can't reliably spot an AI scam by how it looks or sounds. Check the request instead: anything that changes payment details, rushes you, asks for a login or wants you to open a file gets confirmed through a number or channel you already trust. In one bank's test, people told real images from AI ones just 42% of the time.

We read the FBI's 2025 crime report, the Microsoft and Google threat reports and this week's incident notices for this post on October 31, 2026. On November 4 we updated the Binance section from the exchange's own report and the FBI's alert. Binance says its investigation is still going on, so that part may change again.

Photo by Andrea Piacquadio on Pexels

Why the old scam advice stopped working

For years, the best scam filter in business was a misplaced apostrophe. AI has now learned punctuation. The UK's National Cyber Security Centre warned in 2024 that generative AI can write lures without the spelling and grammar mistakes that used to give phishing away. The FBI said the same that December, adding that criminals use AI translation to clean up messages aimed at US victims.

Looking harder doesn't fix it. In research by Australia's CommBank, 89% of people said they could spot an AI-generated scam. When tested, they told real images from AI ones 42% of the time, which is worse than guessing. Among small businesses in the same survey, half of deepfake attempts arrived by email, and only 55% had cross-checked supplier payment details in the past six months.

What the FBI's AI scam numbers show

The FBI's 2025 Internet Crime Report is the first big official count we've seen: 22,364 complaints that mention AI, with $893.3 million in reported losses. Most of that isn't fake emails to companies.

Investment scams make up 71% of the total. Business email compromise, the fake-boss and changed-bank-details fraud that hits small firms, looks tiny by comparison. Overall, that crime cost $3.05 billion across 24,768 complaints, and only 135 of those complaints mention AI.

Read those AI numbers as a minimum. A complaint only counts if it refers to AI, and a well-written email doesn't announce who wrote it. The FBI makes the same point about investment scams: losses passed $8 billion overall, which it says shows many victims don't realize how far AI may be involved.

AI scam examples aimed at small businesses

Fake invoices and changed bank details

ICAEW, the UK accountants' body, describes an Australian construction firm that paid an AU$900,000 invoice in 2024. The email came from the supplier's real address, because the supplier's system had been hacked and the account details swapped. The firm's bank recovered most of it. A lawyer quoted in the same piece says AI now adds convincing logos, layouts and supporting documents if you push back.

It works in your direction too. If a scammer gets hold of your invoice, your client could receive a perfect copy with someone else's bank details. Put your payment details on every invoice you send, and tell regular clients that you'll never change them by email alone. The expense version of this trick is AI fake receipts, which need different checks.

Fake clients, recruiters and job posts

Microsoft reported on October 29 that several North Korean hacking groups now use AI across their operations. Between June 8 and 12, one group posed as recruiters for a made-up AI company, then ran similar outreach through two more front companies, with tailored websites and social profiles to match.

Freelance marketplaces see a smaller version. Upwork's scam guide warns about client profiles with AI-generated headshots and invented work histories, "AI training" jobs that charge an upfront fee, and interviews moved to Telegram or WhatsApp before a contract exists. Real AI training platforms exist too, so check before you judge: Outlier AI is legit, with some caveats. In July, a report from Upwork's Reddit community described a job post asking freelancers to open a compressed file called "Requirements and Budget". Inside, the report says, was a script that downloads and runs malware.

Fake AI apps and extensions

AI tools have become the bait. A campaign reported by The Hacker News uses fake GitHub projects posing as ChatGPT and Claude desktop apps to install crypto miners. On October 29, a malicious copy of the Claude Code extension sat in Cursor's marketplace for about 11 hours, exposing up to 5,000 developers. Cursor says Anthropic didn't make that version, so even a listing with the right name isn't proof.

Download AI apps from the company's own site, and be slow to install anything you found through an ad or a search result. Cisco researchers also found password-stealing malware that sends what it collects to AI models to build a profile of the victim, though they didn't track whether it has been used in real attacks. Malware like that is often hidden in counterfeit software.

Stolen logins

A newly exposed ransomware operation shows how little of this depends on AI. The Register reports that the n0n crew ran up more than $7,216 in AI bills writing and debugging its scripts. Yet in every break-in that researchers at SOCRadar documented, the way in was a weak or stolen password, and no software flaw was exploited. That's hiring a robot locksmith for a door with the key under the mat.

Google's May 2026 report points the same way. The first zero-day exploit it believes was developed with AI bypassed two-factor login, but the attacker still needed a valid username and password first. So the boring fix keeps its value: a different password for every account, from a password generator or manager, plus two-factor login on your email and bank. Just don't ask a chatbot to make them up, because AI-generated passwords are weak for reasons we've covered.

AI voice scams: when the caller sounds right

The largest voice case we found this year hit a bank, which should end any idea that this only catches the unwary. Reuters reported, citing two sources, that Fideuram, the private banking arm of Italy's Intesa Sanpaolo, lost €95 million ($108 million). Its chairman got a WhatsApp message that appeared to come from Intesa's CEO, followed by a call in what sources say was an AI copy of a senior lawyer's voice. More than half was recovered, and about €36 million is still missing.

Photo by Gustavo Fring on Pexels

The raw material is easy to get. In Florida, the Osceola County sheriff says scammers cloned a captain's voice for fake arrest warrant calls. A security expert told the same station that about 20 seconds of audio is enough, and some systems need less than five. If you've posted a podcast, a sales video or a long voicemail greeting, that audio exists.

The FBI's advice is to hang up, look up the person's or company's number yourself, and call it. Its other suggestion is a secret word that only the real person knows. In the CommBank survey, 74% of people said they should set one up and 20% had. Agree one with your business partner, your bookkeeper and anyone else who can move your money.

How to tell if an email is AI generated

Mostly you can't, and it wouldn't settle anything if you could. Plenty of honest clients write with AI, and AI detectors make enough mistakes that a score proves little. What you want to know is whether the request is real. Here's a made-up example of the kind of email to expect:

Nothing in it reads wrong. The job is real, the invoice number is real, and the tone sounds like Dana. But it changes where money goes, sets a deadline and points you to an attachment. Any one of those is enough to pick up the phone and call the number you already have for Dana, not the one in the signature. Then check the sender's full address, the reply-to address and where any link leads, knowing that a hacked account can pass all three.

What the Binance and Polymarket hacks do and don't show

Binance has now published its own initial report, and the first steps match the scams above. In June, an engineer searched online for a popular AI coding tool and installed it with a tampered download. Binance caught that one and replaced the laptop, but the attackers had learned who worked on its wallet systems. In September, another engineer was contacted through a real job-hunting site by someone posing as a project director. A video interview with an "HR representative" came first, and the interviewer shared a personal social media account to build trust. The second round, held in a chat app, ended with an "assignment" to download onto his work computer. It ran malware as he set it up, and Binance's security software didn't catch it.

On October 20, the attackers used that laptop and an internal AI tool to take control of two more engineers' computers. Binance hasn't said how the AI tool was used. Nine days later, they used stolen credentials to reach the service that signs wallet transactions. Nothing Binance has published says AI wrote the lures or ran the theft, and the FBI says North Korea's TraderTraitor group was responsible.

Those first steps would work just as well on a one-person business. A coding test from a new contact is the same trick as the "Requirements and Budget" file above, and an AI tool found through a search result is the same as the fake ChatGPT apps. Binance says it stopped 57 suspected North Korea-linked attempts this year, using methods that include fake job interviews, fake video-call apps and malware disguised as popular developer tools.

Polymarket's loss came through a script from an outside provider, which researchers say worked through logged-in sessions, with no sign that it stole passwords. CNBC reports that the provider was a website monitoring tool, DebugBear, according to a person familiar with the matter. So keep your work laptop clean, and remove plugins and extensions you no longer use. Each one is a piece of someone else's code running where your money and accounts are.

The day after the Binance theft, Chainalysis disclosed its own breach, saying an outsider used "offensive agentic techniques", a phrase that generally points to AI agents. It says the breach is unrelated to Binance and hasn't explained what those techniques were. BleepingComputer reports that customers were told to watch for scammers posing as the company and offering to recover funds. Binance says it will cover all losses and has already replaced most of the stolen assets with its own funds, so treat any stranger offering to get your Binance money back as the next scam.

How to spot AI scams before you pay or click

These checks work however good the fake is, and none of them needs any software:

• New bank details: call a number you already have, every time. Replying to the email only reaches whoever sent it.
• A rush or a secret: slow down. Nobody in the history of business has been told to keep a wire transfer quiet for a good reason.
• A familiar voice or face asking for money: hang up and call back, or ask for your secret word. A convincing voice is no longer evidence.
• A file or link from someone new: ask for the details as plain text in the platform's own messages, and don't run a "test task" on the computer you work on. A real client can type out a brief.
• A move off the platform: stay put until there's a contract. Upwork warns that its payment protection doesn't cover work paid outside the site.
• A new AI tool: get it from the maker's own site rather than a search result, and check the address before you sign in.

If you already paid or clicked

Speed decides how much comes back. The FBI's team for freezing fraudulent transfers handled 3,900 incidents in 2025 and froze $679 million of the $1.16 billion involved, a 58% success rate.

1. Call your bank now. Ask for a recall of the payment, and ask what paperwork it needs.
2. File a report, whatever the amount. In the US, go to ic3.gov with the full transaction details. In the UK, use Report Fraud on 0300 123 2040, and in Scotland call Police Scotland on 101.
3. If you opened a file or entered a password, change your email and bank passwords from a different device, sign out of other sessions and turn on two-factor login.
4. Warn the people involved. Tell your clients and suppliers if your account may have been used, and flag the profile or post on the platform.

UK readers can also forward suspicious emails to report@phishing.gov.uk and texts to 7726, which GOV.UK lists as the official routes.

How we made this: researched and drafted with AI assistance, with any figures checked against their sources and the post reviewed by Hamza before publishing.

Frequently asked questions

Sources