AI phishing in 2026: how to detect a credible message

You receive a message from your CEO. The tone is right. The project it mentions exists. The signature is perfect. It only asks you to review a Microsoft 365 document before 17:00.

The message may be legitimate. It may also have been prepared using public information, written by AI, and sent from a domain almost identical to yours. Good English is no longer enough to tell the difference. You need to examine the request, the channel, and the destination.

These figures do not measure the same thing: ENISA observes intrusion vectors across Europe; Cybermalveillance.gouv.fr counts assistance journeys in France. They therefore mean neither that “60% of SMEs are hacked” nor that “80% of emails are written by AI”.

ENISA also reports that in early 2025, AI-assisted phishing campaigns may have accounted for more than 80% of social engineering activity observed worldwide. The word “assisted” matters: AI may only have been used to translate, rephrase, or personalise. This figure, which ENISA cites from third-party sources, cannot reliably attribute every message to a model.

Read the ENISA 2025 analysis and the Cybermalveillance.gouv.fr 2025 review.

ANSSI describes generative AI as a tool for performance and scaling, not as an autonomous attacker. In February 2026, the agency said it had identified no system capable of carrying out every stage of a cyberattack on its own. People still control the scenario, infrastructure, choice of targets, and exploitation.

Read the ANSSI overview of the generative AI threat.

1. Step 1 — Observe

   Collect public information

   Websites, LinkedIn, press releases, job listings, and exposed documents reveal names, roles, suppliers, tools, and busy periods.

2. Step 2 — Prepare

   Build a credible pretext

   AI summarises the information and adapts the vocabulary to the target: accounting, HR, management, or IT support.

3. Step 3 — Approach

   Send an email, SMS, QR code, or call

   The channel is chosen according to the desired action. A lookalike domain, compromised account, or cloned voice increases trust.

4. Step 4 — Prompt action

   Obtain a login, file, or payment

   The victim is directed to a login page, attachment, change of bank details, or urgent transaction.

5. Step 5 — Exploit

   Reuse the access or money

   The attacker steals a session, reads email, follows up with contacts, deploys malware, or transfers the funds.

• An unusual action: signing in again, installing a tool, sharing a document, buying gift cards, changing bank details, or making a transfer.
• Urgency or secrecy: “before 17:00”, “I'm in a meeting”, “don't tell anyone”. Pressure prevents verification.
• An inconsistent address: the display name is correct, but the full domain differs by one letter, hyphen, or extension.
• A change of channel: a conversation that began by email moves to WhatsApp, SMS, a QR code, or an unusual video call.
• A login initiated by a link: the page looks like Microsoft 365, Google, or DocuSign, but its domain does not belong to the service.
• An MFA factor requested or pushed without any action from you: a six-digit code, approval notification, or device pairing.
• A bypassed procedure: the message specifically asks you to avoid a callback, dual approval, or the usual purchasing process.

Subject: Atlas contract — access the file before 16:00 Hi Leah, the firm has just uploaded the latest version. I'm leaving for a meeting; can you check the figures and confirm on my mobile? The link expires today: sharepoint-revue-docs.com/atlas. Please keep the file confidential until the announcement.

The English is correct and the project may be real. Yet four points require verification:

1. The domain is neither sharepoint.com nor the company's Microsoft 365 domain.
2. The short deadline pushes you to click before thinking.
3. The reply channel changes to a mobile phone.
4. The confidentiality request discourages independent verification.

The right response is not to reply to the email. Open SharePoint from your usual bookmark or call the person on their saved number. If the document exists, you will find it without the link.

These services provide indicators, not an absolute verdict. A link absent from blocklists can be malicious if it was just created. Conversely, an SPF or DKIM failure does not always prove fraud: forwarding or a mailing list can also alter the message.

Before clicking: check a link's reputation

• Google Safe Browsing — Site Status: paste the link address without opening it to see whether Google has already classified it as dangerous. “No data available” does not mean the site is safe.
• VirusTotal — URL or domain search: compares detections from several engines and reputation services. Search for the URL or its domain first. Never upload a contract, invoice, exported email, or internal file: standard submissions may be shared with the service's partners and customers.

Understand where an email came from

• Google Admin Toolbox — Messageheader: turns an email's full headers into a readable route. It helps identify the servers involved and SPF, DKIM, or DMARC results. It cannot decide on its own whether a message is legitimate.
• Microsoft Message Header Analyzer: an alternative suited to Microsoft 365 and Outlook environments. Before pasting a header into an external service, check your internal policy: it contains email addresses, server names, and routing information.

Measure the company's exposure

• Have I Been Pwned: checks whether a work address appears in known data breaches. A positive result does not prove that the account is currently compromised, but it warrants checking the password, MFA, and sessions. The service also offers domain monitoring after ownership has been verified.
• MxToolbox — DMARC, SPF, and DKIM: allows the domain owner to check email protection records. For DKIM, you need to know the selector used by Microsoft 365, Google Workspace, or your provider.

Report an incident or request help in France

• Signal Spam: reports a phishing email or URL without visiting the page. The message's technical data helps organisations that can block the site or identify the sender.
• 17Cyber: a free public diagnosis and assistance service, available 24 hours a day, 7 days a week to individuals, businesses, and local authorities. Depending on the situation, it provides recommendations and may offer a conversation with a police officer or gendarme.

Arup: $25 million lost after a fake video conference

In January 2024, an employee of engineering company Arup in Hong Kong received a message attributed to the CFO. He was initially suspicious. A video conference reassured him: he believed he could see and hear the CFO and several colleagues. Every participant except him was synthetic. The employee made 15 transfers to five accounts, totalling approximately $25 million. The company confirmed the use of fake voices and images.

The lesson is not “learn to spot a deepfake by eye”. A credible video can deceive. The reliable defence would have been independent payment verification and dual authorisation.

Read CNN's account, with confirmation from Arup.

In France: a manager hangs up and blocks the fraud

In a case shared by the DGSI and reported in early 2026, the manager of a French industrial site was contacted by video call by someone who appeared to be the group's CEO. The face and voice were imitated. The caller requested a transfer of funds related to a confidential acquisition. The manager found the approach unusual, ended the call, and contacted management through the usual channels. The attempt failed.

The contrast with Arup is instructive: image quality did not determine the outcome; the verification procedure did.

Read the case and its analysis.

1. Priority 1

   Phishing-resistant authentication

   Enable passkeys or FIDO2/WebAuthn keys as a priority for email, administrator accounts, and finance roles. An SMS or six-digit code is still better than a password alone, but a fake site can relay it in real time.

2. Priority 2

   Out-of-band verification for sensitive actions

   For new bank details, an exceptional transfer, sharing a secret, or installing software, call back on a number already known or request approval in the internal tool. Never use the number in the suspicious message.

3. Priority 3

   SPF, DKIM, and DMARC on the domain

   DMARC in reject mode reduces exact impersonation of your own domain when properly aligned. It blocks neither a lookalike domain nor a compromised partner account: it is one layer, not a guarantee.

4. Priority 4

   A report button and a blame-free response

   Staff must be able to report in one click and alert someone immediately if they clicked. A prompt alert allows you to revoke a session, block a domain, and warn other recipients.

5. Priority 5

   Short, regular exercises

   Test several scenarios: a Microsoft 365 link, QR code, change of bank details, text from the CEO, and MFA notification. Focus on reporting time, not just click rate.

See CISA's MFA recommendations and the ANSSI authentication guide.

• Immediately alert your IT contact or provider, even if nothing unusual appears.
• From a clean device, change the password if you entered it and revoke all active sessions.
• Reject every unsolicited MFA notification and check that no new factor or device has been registered.
• If a payment or bank details are involved, contact the bank immediately and then file a police report.
• Keep the message, headers, URL, timestamps, and screenshots; do not continue browsing the fraudulent site.
• Warn other recipients and look for forwarding or deletion rules added to the mailbox.

• ANSSI — Synthèse de la menace sur l'IA générative face aux attaques informatiques, February 2026: observed offensive uses, limitations, and absence of a fully autonomous attack.
• ENISA — Threat Landscape 2025: analysis of 4,875 incidents in the European Union between July 2024 and June 2025.
• Cybermalveillance.gouv.fr — Menaces visant les professionnels en 2025: statistics from assistance journeys, not from all French companies.
• CISA — More than a Password: hierarchy of MFA methods and recommendation of FIDO/WebAuthn.
• CNN — Arup, victime d'une fraude au deepfake de 25 millions de dollars: facts confirmed by the company and Hong Kong police.
• La Banque Postale — IA et fraude, 2026: French case drawn from a DGSI briefing and analysis of useful controls.

Would you like to assess your teams' responses without tricking or blaming them? Cyber Expert designs role-based simulations and measures reporting time. Arrange an initial discussion.