UTA Researchers Show 10 Posts Can Fuel AI Phishing, While Filters Miss Most Attacks
Updated
Updated · Fox News · Nov 10
UTA Researchers Show 10 Posts Can Fuel AI Phishing, While Filters Miss Most Attacks
3 articles · Updated · Fox News · Nov 10
Summary
University of Texas at Arlington researchers found that as few as 10 public social media posts can give AI enough detail to craft highly personalized phishing emails.
The team analyzed 200 public Instagram accounts and generated more than 18,000 emails, with study participants rating the AI-written lures as more convincing and less suspicious than real-world phishing examples.
Three AI safety tools—Llama Guard, ShieldGemma and WildGuard—caught only 0% to 19% of requests for personalized phishing under default settings, showing major gaps in current safeguards.
After researchers tuned those protections to flag context-aware phishing, detection rates rose to as high as 99% in tests, though the study did not measure real-world click rates or guarantee live blocking.
The findings suggest scammers can weaponize public posts without hacking accounts, increasing pressure on AI companies to harden guardrails and on users to limit public sharing and verify unexpected emails.