GuideBy the PhishSheriff team · June 2025 · 8 min read

Phishing in 2025: What Changed and What to Do About It

A practical look at how phishing changed in 2025: AI-written lures, business email compromise aimed at finance teams, and the speed at which a click becomes a compromise. With the programme changes that follow from each.

Key points

Generative AI has removed many of the tells employees were taught to spot
Business email compromise targets the people who move money, not the mail filter
Credentials can be entered within moments of a click
Reporting speed deserves as much attention as click rate

On this page

Why the threat looks different

Employees are no more or less careful than they were a few years ago. What has changed is the supply side. Generative AI has cut the cost of producing a convincing, well-written lure from hours of manual effort to moments, and attackers use it.

The result is a shift from broad, low-quality campaigns towards targeted, contextually accurate messages sent at volume. An attacker no longer has to choose between scale and quality. Campaigns are increasingly timed around public events, such as earnings announcements, acquisitions and leadership changes, when a themed request feels plausible.

AI-written lures remove the old tells

Classic awareness training taught people to look for spelling errors, generic greetings and implausible stories. AI-written lures often have none of these. They can refer to real tools, copy the tone of a colleague and arrive at a plausible moment in a business process.

For programme owners, the implication is that simulations based on old templates may measure resistance to a threat that has moved on. Exercises should reflect the quality of current attacks, and training should focus on the request (an unusual payment, a new login page, a change of bank details) rather than on surface mistakes.

Business email compromise and finance teams

Business email compromise (BEC) rarely contains a payload: no link, no attachment, nothing for a secure email gateway to inspect. It is a social-engineering request aimed at the people who move money, such as treasury, accounts payable and finance controllers.

Two factors make it harder to recover from. Faster payment rails shorten the window for recalling a fraudulent transfer, and attackers increasingly reinforce the email with a phone call from someone claiming to be a senior executive. Organisations that rehearse only the email half of the pattern are practising against half of the attack.

Compromise happens fast

When a credential-harvesting page works, the user often enters their details almost immediately after clicking. Mobile users are especially exposed: small screens hide address bars and security indicators, and mobile mail clients often truncate sender addresses.

This affects response planning. If a playbook assumes there are minutes or hours between a click and a compromise, it may be too optimistic. Fast reporting matters because it gives responders a chance to reset credentials and contain a mailbox before it is used.

Programmes that change behaviour

Programmes tend to show lasting improvement when they run regularly rather than once a year, adjust difficulty to the people taking part, and cover the channels attackers actually use. One-off annual training is easy to complete and easy to forget.

Expect improvement to be uneven. Early gains usually come from removing careless mistakes; later gains depend on habits such as pausing, verifying through a known contact and reporting, which take longer to form.

What to do next

Three changes follow. First, retire static template libraries: if simulations are easier than real attacks, the measured click rate overstates your resilience. Second, extend exercises beyond email to the voice, SMS and QR requests that often form the second stage of an attack. Third, measure reporting as well as clicking, including how quickly suspicious messages reach the security team.

Finally, pair the people programme with the domain layer. DMARC enforcement helps receiving systems reject mail that falsely claims to come from your domain, so fewer impersonation attempts reach employees at all. It does not stop lookalike domains, display-name spoofing or messages from compromised accounts, which is why the human programme still matters.