AI-generated spearphishing: What a participant study shows
Personalised deceptive messages can be automated. Experimental click rates must be distinguished from account compromise and universal success.
Révision éditoriale :
Le texte de cet article est disponible en anglais dans cette version du site.
Contexte et sources
Evaluating Large Language Models' Capability to Launch Fully Automated Spear Phishing Campaigns compares email groups with 101 participants in total. Human-expert and automated emails achieved comparable click rates in that setup. Clicking was the measured action, not a completed account takeover. Findings depend on participants, messages and study conditions. [1]
Portée et limites
The sample does not represent every workforce. Clicking a link does not itself establish successful data theft. [1]
En termes simples
A hypothetical fraudulent message may read like normal business correspondence. Its writing quality does not establish that the apparent sender actually authorised the request.
Recommandations pratiques
We recommend teaching verifiable procedures rather than relying on spelling mistakes as warning signs. Confirm payment changes and unusual information requests through known channels. Make internal reporting straightforward. For simulations, define objectives and privacy safeguards beforehand, and assess reporting and containment alongside clicks.