AI Social Engineering: Attacks Now Span Every Channel

AI has changed cyber risk in two ways: precision and reach. Attacks now match your role, tone, and timing. They also move across email, chat, and video inside one conversation. Spelling errors no longer give them away. Defenders need new habits: check the context, verify through a second channel, and report fast.

The Old Warning Sign Is Gone

For years, training taught one simple tell: bad grammar. Scam emails had typos, odd phrasing, and clumsy layouts, so staff learned to look for them. Then AI removed that tell, and AI social engineering now reads like a real colleague wrote it. A model now writes clean, natural prose in seconds, and it copies the tone of a real colleague.

The results are measurable. In a 2024 IEEE Access experiment, phishing emails written by GPT-4 drew 30 to 44% click-through. Generic phishing drew 19 to 28%. So polish is no longer proof of anything.

Attackers also scale. One person can now produce thousands of tailored messages in an afternoon. Each one can mention a real project, a real manager, or a real deadline. That is why phishing attacks stay so successful, even against careful staff.

Attacks Now Move Across Channels

Old phishing was a single email. Modern attacks are conversations. An attack might start with an email that looks routine. Then a chat message follows, asking for a quick favor. Finally, a “verification” call arrives on a Friday afternoon, when people want to finish and leave.

Each step borrows trust from the last. A person who ignores the first email may act on the third contact, because it feels like a real exchange. Meanwhile, most security tools watch one channel at a time. The email gateway sees the email. The chat platform sees the chat. Nobody sees the whole chain except the target.

That makes people the only sensor that spans every channel. It also explains why assessing human behavior across email, calls, and social platforms matters more than email tests alone.

A Deepfake Video Call Cost $25.6 Million

The Arup case shows where this leads. In January 2024, a finance employee at the engineering firm’s Hong Kong office received a message about a confidential transaction. It appeared to come from the UK-based chief financial officer. The employee was suspicious at first.

Then the attackers moved the conversation to a video call. On screen were the CFO and several colleagues. All of them were deepfakes. Reassured by familiar faces and voices, the employee made 15 transfers worth about HK$200 million, or US$25.6 million, to five bank accounts.

The fraud came to light only when the employee checked with the company’s UK head office. Arup later confirmed to Fortune that fake voices and images were used, and that its systems were not breached. Nothing was hacked. A person was persuaded. This is the pattern behind modern deepfake scams.

Context Beats Content

If wording no longer helps against deepfake and multichannel attacks, what does? Context does. Instead of asking “does this message look right?”, ask “does this request fit?” Three questions help:

• Does this request match this person’s normal behavior? A CFO who never messages finance staff directly is a warning.
• Is the channel normal? A sudden move from email to a new chat app or a private call deserves a pause.
• Does it involve money, access, or secrecy? These are the goals of nearly every attack.

Watch for a mix of urgency and secrecy, since real business rarely needs both. When you see it, stop and verify. Changed bank details are another classic trigger. Teams that handle payments should know how invoice fraud and fake vendor scams work, because these attacks reach finance staff first.

Verification Rules That Work Against Deepfakes

Do not rely on spotting a fake face. Detection tools lag behind generators, and people are poor at it. Rely on process instead. A good rule works even when the fake is perfect.

• Call back on a known number. Use the number in your directory, never the one in the message.
• Require two approvers above a set amount. No single person should release a large payment.
• Hold new payees for 24 hours. Delay gives the real owner time to object.
• Never approve money by chat or video alone. Treat a call as a request, not as proof.
• Escalate anything both urgent and secret. Make it normal to ask a second person.

These rules take minutes to write and cost almost nothing. However, they only work if leaders back them. If a senior executive pushes staff to skip a step, the rule must still win. They also protect against CEO fraud in every form, with or without AI.

Train for Every Channel

Training that covers only email leaves multichannel attacks a way in. Attackers already use text messages, chat apps, and cloned voices, so staff should practice against each of them in a safe setting.

Start with the channels where your money and data move. For example, a good drill might begin with an email, follow with a chat message, and end with a call that asks for a payment. Staff then practice spotting the chain, not just one message. Simulate SMS and chat lures, and then add voice, because a cloned executive voice tests a different reflex than a suspicious link does. A voice phishing simulation with cloned voices lets teams rehearse the callback rule before it matters.

Measure what people do next. The best sign of progress is reporting speed, not a lower click rate. Praise near misses, so staff feel safe to speak up.

Govern Your Own AI Agents Too

Attackers can trick software as well as people. An AI agent that reads email, moves files, or books payments can be steered by hidden text. So treat each agent like a new colleague with limited trust.

Give every agent its own credentials, and keep its access narrow. Require a human sign-off for high-risk actions such as moving money or changing permissions. Log what the agent does, so you can trace a mistake later. In short, apply the same verification rules to machines that you apply to people.

The Bottom Line

AI social engineering did not create new goals for attackers. It made the old ones faster, more personal, and harder to see. The answer is not better eyesight. It is better habits: check the context, verify through a second channel, and report at once when something feels off.

Frequently Asked Questions

Anjali Chauhan

Anjali is the Cybersecurity Manager at Kratikal, leading a team focused on strengthening security through rigorous vulnerability assessments and penetration testing. With expertise across web, network, and cloud environments, she drives strategies to safeguard clients’ critical assets while mentoring her team and staying ahead of escalating cyber threats.

Anjali Chauhan

Anjali is the Cybersecurity Manager at Kratikal, leading a team focused on strengthening security through rigorous vulnerability assessments and penetration testing. With expertise across web, network, and cloud environments, she drives strategies to safeguard clients’ critical assets while mentoring her team and staying ahead of escalating cyber threats.