Index
- No more warning signs
- Context creates trust
- Seeing isn't believing
- Many channels, one attack
- What to do now
- Conclusion
AI and social engineering: the new attack surface
Phishing without warning signs
For years, employees were taught to spot phishing by its flaws: clumsy wording, spelling mistakes, a logo that didn't look quite right. Generative AI has removed those flaws. Language models write flawlessly in German, English and dozens of other languages, and match a company's tone of voice. Convincing emails and copies of real login pages can be produced in minutes. The signals people were trained to look for are disappearing, and what remains often looks exactly like a legitimate message.
Context creates trust
Polished language alone is not what makes these attacks work. Attackers combine it with publicly available information (OSINT) such as LinkedIn profiles and press releases. This lets them tailor every message to a single person, with real project names, real colleagues and a plausible reason to act. A message that mentions the right supplier, the right deadline and the right manager doesn't raise suspicion. It feels like part of a normal working day. And with AI, this level of personalisation no longer takes days of research per target. It scales.
When seeing and hearing is no longer believing
Faced with a suspicious email, many employees have learned to do the right thing: ask for a quick call or a video meeting. For a long time, hearing or seeing the person was proof enough. AI has broken this assumption.
A voice clone often needs nothing more than short audio recordings, for example from an interview or a podcast. Many executives have hours of such material publicly available. Deepfakes in live video calls are no longer a vision of the future either, and they can be created with freely available tools. If the employee calls the number in the email, or the attacker sets up the meeting, the check ends up vouching for the fraud. The channel that was supposed to verify a request can now be part of the attack.
One attack, many channels
The most dangerous attacks don't stay in one channel. An email announces a request, a text message adds urgency, and a call in a familiar voice confirms it. On its own, any one of these messages might make a trained employee pause. Together, they answer each other's doubts. The attack is prepared in advance and aimed at one person, with a story that holds together across every channel.
This is where email-only training reaches its limits. Someone who has learned to question a perfect email may still trust a perfect voice, simply because they have never experienced a fake one.
What organisations should do now
Train for the attacks that actually happen. Awareness programmes should include realistic voice and video scenarios, not just generic phishing emails. Employees who have experienced a cloned voice in a safe training environment are far more likely to pause when it happens for real.
Make verification a routine, not an exception. Payment approvals, password resets or changes to bank details should never be based on a call, a video meeting or an email alone. Call-backs to known numbers, the four-eyes principle or agreed code words are simple and effective measures, and they work even when the voice sounds perfect.
Measure the reporting rate, not the click rate. A single click is rarely the end of the story. What matters is how quickly a suspicious call or message is reported. A high reporting rate gives the security team valuable time and significantly reduces risk. This requires a simple reporting process and a culture in which nobody is afraid to admit a mistake.
Combine people and technology. It is the security team's job to protect employees with appropriate technical measures, such as phishing-resistant multi-factor authentication and clear access concepts. Training complements technical controls, but it does not replace them.
Conclusion
AI has made social engineering more convincing, more scalable and harder to detect, and a familiar voice is no longer proof of identity. Organisations should therefore rely less on spotting individual warning signs and more on fast reporting, clear verification processes and realistic preparation. At revel8, this is why we train teams with OSINT-personalised simulations across email, text messages, phone and video, so that the first real deepfake call isn't also the first one anyone has ever experienced.