You posted photos from a trip to Brazil on social media. The next day an email arrives from the hotel where you stayed: thanks for the visit and a discount on your next stay. The email mentions your trip, the hotel name looks right, the link seems reliable. But the hotel sent nothing. The email was generated by a neural network that gathered details from your public posts.
Researchers at the University of Texas at Arlington (USA) showed that widely available artificial intelligence (AI) tools can collect information from public social media posts and quickly create mass personalized scam emails. This technique is called context-dependent spear phishing: an attacker approaches a specific person using details of their life to get them to click a malicious link and give up passwords or bank card data.
"Anyone can attack, and that is frightening," says Shirin Nilizadeh, an associate professor in the UTA Department of Computer Science and Engineering who led the work. According to her, with such models attackers do not need to do much: it is enough to say what is required.
Spear phishing is not new. Nilizadeh notes that previously most targeted attacks relied on static data - a job title or place of work. The new work studies a more sophisticated variant that uses details of daily life posted online. "Now scammers can take people's social interactions and everyday life and insert them into emails," she says. "It is more personalized and can make people more vulnerable."
Study co-author and graduate student Elham Pourrabbas Vafa explains that AI tools pulled details from recent posts, including photographs, and used them to compose emails. The researchers tested emails on topics of travel, events people were interested in (such as concerts), and even concerns they shared online. The email subject, sender name and links were tailored to each recipient. "A person will find such an email very plausible because it mentions details of their recent experience," says Pourrabbas Vafa.
In a study involving users, the AI-generated emails received high quality ratings - which, according to the authors, is what makes the attacks dangerous.
Companies developing AI already use safety filters intended to block harmful requests. However, the researchers found that such filters mostly did not recognize requests to create personalized scam emails. The team tested several common tools, including Llama Guard, ShieldGemma and WildGuard. With default settings they detected only 0% to 19% of attacking requests. When the filters were given specific instructions to recognize such attacks, the detection rate rose to 99%.
"We call on companies working with AI to take such attacks seriously," says Nilizadeh. "They need to implement safeguards so that their tools are not used to create high-precision social engineering attacks."
The authors did not include the exact instructions used to create the scam emails in the public version of the work so that others could not abuse them.
The study has been accepted for presentation at the Network and Distributed System Security Symposium in Seoul (South Korea) in March 2027. Pourrabbas Vafa also presented the work at the Texas Security Seminar at the University of Texas at Dallas.