AI models can use public social-media details to make phishing emails feel personal, researchers at the University of Texas at Arlington say. Their reported tests also found that several tools designed to flag harmful AI requests caught few of the personalised-phishing prompts at standard settings.
Watch Desk analysis
What happened
The researchers used widely available AI models to gather details from people’s public posts and create targeted messages, according to Digital Trends. The messages could refer to recent activities or plans, and the models could adjust sender names, subject lines and links to make emails look more convincing.
The report says participants rated the AI-generated emails highly for quality. That does not show how many people would fall for them, but it does underline why a message that sounds unusually well-informed deserves a second look. The research is due to be presented at a cybersecurity conference in March 2027.
Digital Trends reports that Llama Guard, ShieldGemma and WildGuard flagged only 0% to 19% of the tested requests at their standard settings. With clearer instructions, the detection rate rose as high as 99%, the researchers found.
Why it matters
Personalised phishing has always relied on knowing something about its target. The reported development is that AI can help collect and use those details at scale, making tailored messages less laborious to produce. An Instagram post about a trip or event can become useful context for a stranger trying to make a scam feel familiar.
The findings also raise a practical question for AI safety tools: their performance may depend on how a risky request is framed. A low detection rate in these tests is not a measure of how often scams succeed, and the reported improvement is not a guarantee that clearer instructions will catch every harmful request.
Our read
The useful warning is not “never post anything online”, which would leave the internet looking rather empty. It is that personal details can give a suspicious message an unnervingly plausible opening. Check the sender and destination independently, even when an email mentions something you genuinely did.
For developers, the reported difference between standard and more specific instructions suggests that safety tools need testing against realistic, personalised misuse, not just obvious bad prompts. These are research findings reported ahead of a conference presentation, not evidence of a measured surge in successful scams.
What to watch
- The researchers’ conference presentation and any fuller account of the methods and results.
- Whether independent teams reproduce the findings across models and safety tools.
- Whether safety systems can reliably recognise personalised phishing requests without blocking legitimate uses.
Discussion spark: Should AI developers be responsible for making phishing requests harder to carry out, or does the practical burden belong mainly with users and email providers?
Sources and evidence
- Source update (4 November 2026, 14:50 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.