Study finds public social posts can enable personalized AI phishing

The study examines attacks built from details of people’s everyday lives shared online.[1] The researchers used recent social media posts, including photos, to create personalized emails.[1] They tested messages tailored to recipients’ travel, concert interests and concerns they had shared online.[1] The paper’s authors report that they used Instagram posts from 200 people and that as few as 10 posts per person could produce highly contextualized emails.[2] The findings focus on how online details can shape a message’s relevance to its recipient.[1]

The researchers customized each recipient’s email subject line, sender name and links.[1] The emails were designed to lure recipients into clicking malicious links and sharing sensitive information, including passwords or credit card numbers.[1] In a user study, participants rated the AI-generated emails highly for quality.[1] The paper reports generating about 18,000 phishing emails with five generative AI models and comparing them with 4,000 real-world samples.[2] It also reports that the generated attacks outperformed APWG eCrimeX emails across eight dimensions and elicited lower suspicion among human recipients.[2]

The researchers found that existing AI safety filters largely failed to detect requests to create personalized scam emails.[1] With default settings, the tools detected between 0% and 19% of attack requests.[1] Giving the safety tools specific instructions to recognize these attacks raised detection rates to as high as 99%.[1] The researchers urged AI companies to add safeguards against misuse of their tools for targeted social engineering.[1] For teams assessing AI safeguards, the contrast makes it useful to consider how systems perform both with default settings and with explicit detection instructions.[1]

The researchers left the exact scam-email instructions out of the public paper to reduce the risk of misuse.[1] The paper assesses seven phishing strategies: baiting, scareware, honey trap, tailgating, impersonation, quid pro quo and personalized emotional exploitation.[2] The study was accepted for presentation at the Network and Distributed System Security Symposium in Seoul in March 2027.[1] The findings describe tested approaches and safety-filter results; they do not establish that every public post can be used to create an effective scam email.[1]