We’ve all received those suspicious emails claiming we’ve won something, missed a delivery, or need to urgently update our bank details. Most are easy enough to ignore, especially when they’re full of strange wording or have absolutely nothing to do with us. But what happens when a scam email actually knows what you’ve been up to lately? That’s where things get a little worrying.
Researchers at the University of Texas at Arlington have found that AI can use information people share publicly on social media to create phishing emails that feel personal. And we’re not just talking about knowing your name or where you work. These emails can reference things you’ve recently done, places you’ve visited, or even something you’ve been looking forward to. The researchers showed how widely available AI models could gather these details and turn them into targeted messages, potentially letting scammers reach large numbers of people without spending hours researching each person individually.
That harmless Instagram post could become someone else’s bait
Think about how much we casually share online. Maybe you’ve posted pictures from a weekend getaway, mentioned tickets to an upcoming concert, or complained about a service that hasn’t been working properly. To your friends, these are just everyday updates. To someone planning a scam, however, they’re useful details. The researchers found that AI could pick up on these details, including information from photos, and use them to make fraudulent emails appear relevant to someone’s life. It could also adjust the sender’s name, email subject, and links to make the whole thing look more convincing.
This is a form of spear phishing, where scammers target people with messages designed around their personal information. The difference here is how much work AI can do. The team also asked participants to evaluate the AI-generated emails, which received high-quality ratings. That doesn’t necessarily mean everyone would fall for them, but it does show how convincing these messages can appear.
The tools meant to stop AI scams have a problem of their own
What’s particularly concerning is that AI already has safety systems designed to prevent this kind of misuse. Unfortunately, the researchers found that several of these systems struggled to recognize what was happening. They tested Llama Guard, ShieldGemma, and WildGuard, three tools used to identify potentially harmful AI requests. At their standard settings, they flagged only 0% to 19% of requests involving personalized phishing attacks. Fortunately, the researchers also found a possible way to improve things. By giving these safety systems clearer instructions on what to look for, they pushed detection rates as high as 99%.
The findings, which will be presented at a cybersecurity conference in March 2027, highlight a growing problem with AI-powered scams. The technology isn’t necessarily inventing information about you; it’s using better information you’ve already shared. And that’s probably the biggest takeaway here. Just because an email mentions something personal doesn’t mean it came from someone you trust. If anything, that’s another reason to look twice before clicking a link.
