AI social engineering awareness
Generative AI has made social engineering dramatically more convincing. The grammar is perfect, the voice on the phone sounds like your CFO, and the spear phish references a project you really are working on. The defensive instincts that used to work — spotting typos, sensing a “off” tone — no longer hold. This interactive guide walks through the main AI-enhanced attack patterns, the signals that still give them away, and the process-based defenses that work even when the content is flawless.
How it works
Pick an attack type — AI-generated phishing, voice-cloning (vishing), or deepfake impersonation — and the guide shows the way the attack typically unfolds, the detection signals that remain reliable, and the recommended response. The through line is the same across all three: because AI can fake the content, your defense has to rest on process — out-of-band verification, code words, and approval workflows the attacker cannot reach. Nothing is sent anywhere; the guide runs entirely in your browser.
Building organisational defenses
- Out-of-band verification. Confirm any money, credential, or data request through a separate known channel — never a number or link in the message.
- Approval workflows. Require two-person approval for high-value transfers so no single convincing message can move funds.
- Code words. Agree a shared word for sensitive voice/video requests; a clone will not know it.
- Assume the content is perfect. Train people that flawless grammar and a familiar voice are no longer reassurance — process is.
Understanding each attack type in depth
AI-generated spear phishing
Traditional phishing was sent at volume with generic content and caught by its tells: poor grammar, impersonal salutation, implausible sender. AI-generated spear phishing uses scraped LinkedIn, news, and social media data to personalise at scale. An attacker can automatically generate a message that references your actual manager, your real current project, a real public event in your industry, and addresses you by name — all without human involvement.
The defensive posture has to shift from content inspection (“does this look suspicious?”) to process: any request involving money, credential sharing, account changes, or data access is verified through a known-good channel regardless of how legitimate the message looks. The content is no longer a reliable signal.
Voice cloning (vishing)
Voice-cloning technology can produce a convincing copy of a known voice from a few seconds of audio — available from voicemails, video recordings, conference calls, or social media clips. The resulting audio can be used in real-time phone calls or as a pre-recorded message. The psychological impact of hearing a trusted voice is strong enough that people override doubt they might otherwise act on.
The defense is procedural: any unusual voice request — particularly one involving urgency, money, or credentials — is confirmed by calling the person back on a number you already have in your contacts, not a number given in the call. A shared code word known only to the real person and asked in every unusual call is a simple, reliable secondary check.
Deepfake video impersonation
Real-time deepfake video has been used in documented cases to impersonate executives on video calls and authorise large financial transfers. The attacker joins a call with a video feed that has been overlaid with the target person’s face, generated in real time. The voice may be cloned simultaneously.
The detection challenge is that real-time deepfakes have visible artefacts — lighting inconsistencies, unnatural movement at face edges, lag between lip movement and audio — but these require attention to notice in a normal working environment. Organisational defenses include: requiring specific unexpected physical actions (“please pick up something from your desk”) that a deepfake cannot seamlessly execute in real time, having a code word agreed out-of-band before calls where high-value decisions will be made, and establishing a policy that high-value approvals cannot happen on a video call alone — they require a follow-up in a second channel.
Training people to resist high-pressure manipulation
Social engineering attacks exploit urgency. “The transfer must happen in the next 30 minutes or the deal collapses” is designed to short-circuit the verification reflex. Training that helps people recognise this pressure pattern — and understand that the right response to unusual urgency is always to slow down and verify, not to act faster — is more protective than training focused on spotting technical tells in the content.