Hunting for spelling mistakes to spot a fraudulent email no longer works. Official European reports from 2025 and 2026 agree that criminals are using generative AI to write personalized messages in multiple languages at scale. What a language model cannot fake is everything outside the text: the real sender address, the domain a link points to, and the channel through which you are asked to act. That is where you need to look.
- Phishing dominates: ENISA’s 2026 Threat Landscape attributes 77.8% of recorded social engineering techniques to phishing, with groups using language models to draft tailored messages at scale.
- Grammar is dead as a filter: Europol’s IOCTA 2026 notes that generative AI lets networks operating outside the EU personalize narratives easily and attack European citizens at high speed and scale.
- Verify out of band: European financial supervisors advise confirming any request through a known number or a trusted channel rather than replying to the message itself.
What the official reports actually say
The change is not in the number of scams but in their quality. An email that once gave itself away through clumsy translation now arrives in impeccable English, with the tone of your bank and, sometimes, with your name and your company’s name.
ENISA’s Threat Landscape 2025, covering July 2024 to June 2025, found phishing was the leading intrusion vector, accounting for close to 60% of cases. It cites, with a reported caveat, that by early 2025 AI-supported campaigns made up more than 80% of observed social engineering activity worldwide. The 2026 edition, published in September 2026 with 2025 data, puts phishing at 77.8% of recorded social engineering techniques and notes that groups use language models to draft bespoke messages at scale.
One nuance is worth keeping. The 80% figure in the 2025 report is attributed with a reported qualifier and measures social engineering activity, not emails received by individuals. In the 2026 report, the agency itself tones things down: attackers mainly use AI to improve what they already knew how to do, not to gain new capabilities.
Europol shifts the focus elsewhere, and it is arguably the most relevant point for European readers: networks operating from outside the EU. For years, a scam written from another country was noticeable in the grammar. INCIBE notes that criminals exploit AI’s ability to create credible text in several languages, and with that, the clue disappears.
INCIBE’s 2025 balance, published on 9 February 2026, reports 122,223 incidents managed in 2025, up 26%. Within online fraud, 25,133 were phishing cases, the most frequent method, and 28% of queries to its 017 helpline were phishing, vishing or smishing attempts.
If the text is perfect, what do you look at?
Everything that is not text. European financial supervisory authorities summarize it in their consumer guide to scams in the AI era: they still point to grammatical errors in official-looking documents, but warn these are increasingly difficult to detect thanks to AI. The signals that hold up are different.
1. The real sender address. The display name you see (“Customer Service”, “Tax Agency”) is written by whoever sends the message. The full address is not. Always expand it. A serious service uses its own domains: if the message comes from a mailbox such as @gmail.com or @outlook.com, be suspicious.
2. The exact domain of the link. Hover over it or long-press without opening. The European guide advises checking for extra or missing letters in the company name. If in doubt, free tools can check whether a link is safe before you open it.
3. Urgency. Account blocked, fine, parcel held, refund expiring. European financial authorities describe this as pressure created deliberately so you act before you think, and their advice is direct: if you have doubts, even slight ones, do not act.
4. What they ask for. No legitimate company will ask for your PIN, password, online banking key or security codes by email, SMS, social media or phone, the same guide recalls.
5. The attachment. An executable, a compressed file or an Office document with macros (.docm, .xlsm) from someone you did not expect. Attachment types are a common infection route, and major providers block many of them.
6. The timing. ENISA documents that phishing is seasonal: in at least nine EU countries, tax agencies were impersonated with fake refunds and payment notices. Be more wary than usual during tax season.
Why two-factor authentication is no longer enough on its own
Because phishing kits have evolved too. ENISA reports that Tycoon 2FA was, reportedly, the most observed phishing kit worldwide in 2025. The agency describes it as an adversary-in-the-middle kit designed to bypass multi-factor authentication: the fake page sits between you and the real site, and whatever you enter, password and code included, serves the attacker to log in in real time.
This does not mean two-step verification is useless. It means a code you type can be stolen if you type it on the wrong site. Passkeys do not have that problem, because they are bound to the real domain and there is nothing to type. If a service offers them, enable them.
Always verify through another channel
The rule that sums all of the above is boring but effective: do not reply to the message, leave it. Open your bank’s app, type the web address yourself or call the number on your card. The European guide frames it the same way, even for messages that appear to come from a relative: confirm through a known number and a trusted channel. Many classic signals remain valid; what has fallen is grammar as a filter.
And if you have already taken the bait, change your password from another device, notify your bank if you gave payment details and call 017, INCIBE’s Cybersecurity Help Line.
The uncomfortable conclusion is that we have spent years teaching people to spot phishing by how it is written, and that advice has expired. Repeating it is dangerous because it creates a false sense of security just when the emails are better than ever. The only useful question is no longer whether a message looks authentic, but whether you have independently verified that it is. Fortunately, that does not depend on how well a machine writes.