---
type: "post"
id: "turning-your-social-media-feed-into-bait-a86166"
slug: "turning-your-social-media-feed-into-bait-a86166"
title: "Turning your social media feed into bait."
description: "Researchers at The University of Texas at Arlington say widely available AI tools can turn details from public social-media posts into highly personalized phishing emails."
retrieval_nugget: "Researchers at The University of Texas at Arlington say widely available AI tools can turn details from public social-media posts into highly personalized phishing emails."
published_at: "2026-11-12"
updated_at: "2026-11-12"
record_date: "2026-11-12"
date_kind: "published_at"
topics: ["security"]
entities: ["The University of Texas at Arlington","Shirin Nilizadeh","Elham Pourabbas Vafa","Llama Guard","ShieldGemma","WildGuard","Network and Distributed System Security Symposium"]
related_patterns: ["agent-security-moves-to-runtime-boundaries"]
novelty_label: "new_fact"
what_changed: "UTA researchers report that widely available AI tools can use people’s public social posts and photos to generate personalized phishing emails at scale."
prior_refs: ["https://newruntime.com/posts/barracuda-found-phishing-campaign-targeting-humans-and-ai-d0fd9d/","https://newruntime.com/posts/phishing-emails-are-baiting-your-ai-assistant-analysis-71783e/"]
source_url: "https://www.uta.edu/news/news-releases/2026/11/02/your-social-media-posts-could-fuel-the-next-ai-scam"
source_urls: ["https://www.uta.edu/news/news-releases/2026/11/02/your-social-media-posts-could-fuel-the-next-ai-scam"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:turning-your-social-media-feed-into-bait-a86166"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/turning-your-social-media-feed-into-bait-a86166/","markdown":"https://newruntime.com/posts/turning-your-social-media-feed-into-bait-a86166.md","json":"https://newruntime.com/posts/turning-your-social-media-feed-into-bait-a86166.json"}
markdown_url: "https://newruntime.com/posts/turning-your-social-media-feed-into-bait-a86166.md"
---

# Turning your social media feed into bait.

## Retrieval answer

Researchers at The University of Texas at Arlington say widely available AI tools can turn details from public social-media posts into highly personalized phishing emails.

## Summary
Researchers at The University of Texas at Arlington say widely available AI tools can turn details from public social-media posts into highly personalized phishing emails. In tests, three safety tools detected only 0% to 19% of attack requests with default settings, but detection rose as high as 99% after researchers gave them specific instructions. The gap matters to teams relying on generic filters to contain social-engineering risks.

## What happened
The University of Texas at Arlington announced a study of context-aware spear phishing, an approach that uses details about a target’s everyday life rather than relying only on static information such as job title or employer. The researchers describe attackers using publicly available social-media posts, including photos, to learn about recent experiences and incorporate those details into scam emails. The announcement names Shirin Nilizadeh, an associate professor in UTA’s Department of Computer Science and Engineering, as the study’s lead, and graduate student Elham Pourabbas Vafa as a co-author.

The team tested emails based on travel, events people were interested in, such as concerts, and concerns people had shared online. It says the AI tools could tailor not just the message but also the subject line, sender name and links to each recipient. One example in the announcement imagines someone posting about a trip to Brazil and then receiving a purported hotel email that refers to the trip and offers a discount. UTA says the hotel did not send that message; the example illustrates how familiar context could make a fraudulent request look credible. The researchers report that participants in a user study gave AI-generated emails high marks for quality.

The study also assessed safeguards from Llama Guard, ShieldGemma and WildGuard. With their default settings, these tools detected between 0% and 19% of the attack requests. The researchers then gave the safety tools specific instructions to recognize this class of attack. Under that condition, detection rates rose to as high as 99%. Those figures describe the study’s tests; the announcement does not present them as a real-world prevention rate or say that every product configuration reached the maximum.

The researchers proposed ways to prevent the emails from being generated in the first place and urged AI companies to add safeguards against targeted social engineering. They omitted the exact instructions used to produce the scam emails from the public version of their paper, saying this was to prevent misuse. UTA says the study was accepted for presentation at the Network and Distributed System Security Symposium in Seoul in March 2027. Pourabbas Vafa also presented the research at the Texas Security Workshop at The University of Texas at Dallas. The announcement does not specify a product release, price, or general availability change.

## What the archive adds
New Runtime’s earlier coverage described an email assistant’s security as depending on the boundary between untrusted message content and the actions or recommendations its runtime can produce ([Phishing emails are baiting your AI assistant](https://newruntime.com/posts/phishing-emails-are-baiting-your-ai-assistant-analysis-71783e/), October 22, 2026). It also covered Barracuda’s report of a campaign targeting both employees and AI inbox summarizers ([Barracuda found phishing campaign targeting humans and AI summarizers in the same email](https://newruntime.com/posts/barracuda-found-phishing-campaign-targeting-humans-and-ai-d0fd9d/), October 30, 2026). UTA adds a research finding about generating personalized attacks and the tested detection gap, rather than a new incident.

## Trend signal
This reinforces the runtime-boundary security pattern: [agent security moves to runtime boundaries](https://newruntime.com/patterns/agent-security-moves-to-runtime-boundaries/). The finding is not a direct test of agent runtimes, but it shows why safety cannot rest on a broad instruction or assumed model behavior alone. Teams need controls that identify risky inputs and constrain what systems can do with personal context.

## Why now
UTA says the work was accepted for presentation at NDSS in March 2027 and had also been presented at the Texas Security Workshop. The announcement frames the research as a warning about tools that are already widely available, rather than a hypothetical capability tied to a future product launch.

## Connections to other events
The study meets New Runtime’s earlier coverage of phishing aimed at both people and AI summarizers. In both cases, attackers exploit the way AI systems process ordinary communications; UTA’s finding adds a different angle, showing how public social details can be used to make an initial message feel personally credible.

## Why it matters
For teams building AI products, the result is a reminder that model safety tests should cover the whole path from information gathering to generated output. A filter that handles generic harmful requests may still miss a request whose ingredients look like ordinary public facts but whose purpose is targeted deception. The 0%–19% default-setting result is a warning against treating a safety component’s presence as evidence of effective coverage.

The 99% maximum after explicit instructions is encouraging, but it is not a turnkey defense. Teams should test their own system configurations, inputs and thresholds, and should avoid making personal context available to generation steps unless the use is justified. Products that summarize or act on messages also need to treat external content as untrusted and limit the actions it can trigger. For users and organizations, public information can lower the effort needed to make a scam persuasive; awareness and independent verification of unexpected requests remain important.

## Limits and uncertainty
UTA’s announcement does not provide the full paper, sample sizes, exact evaluation protocol, confidence intervals, or per-tool detection results. It does not establish how often these attacks succeed outside the user study, how the tools perform against variants of the tested requests, or whether the higher detection rate comes with false positives. The exact generation instructions were withheld from the public paper. The reported rates therefore should not be read as a general measure of protection across models, products, or live attacks.

## Skill to build now
Run a small, controlled safety evaluation against your own system: include requests that combine public personal details with a request to impersonate a trusted organization, and check whether the system blocks or safely redirects them. Record both missed attacks and benign requests incorrectly blocked. Do not reproduce operational phishing prompts in shared test materials; UTA says it withheld its exact instructions to reduce misuse.

## Reader takeaway
Personal details that look harmless in public can make a fraudulent message more convincing when AI helps assemble them. The practical lesson for builders is to test safety controls against targeted social-engineering scenarios explicitly—and verify what happens in their own configuration instead of assuming default filters are enough.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Security](https://newruntime.com/topics/security.md): Explore the security topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)