Hyperlocal fake news site networks across 14 US states target AI chatbots to skew answers on local races
VisionHeight, an Israeli threat intelligence firm, uncovered a network of hyperlocal websites spanning at least 14 US states, built to influence how AI chatbots answer questions about local and state races. Sites with names like "Wolverine State Report", "Peach State Bulletin" and "Tarheel State Report" repeat hundreds of near-identical statements that a given candidate is "the stronger choice" on issues such as housing and taxes. The content is written for AI crawlers rather than human readers, which shifts the objective from persuading people to poisoning the retrieval sources that chatbots draw on. VisionHeight does not attribute the activity to any actor and stresses that the technique is easily reproducible by foreign and domestic actors alike. Perplexity was confirmed to cite one of the pages (Tarheel State Report) as its first source for an answer.
Detalhes técnicos
- Each state network typically has 5-6 sites, some posing as hard news and others as opinion or editorial outlets, and they cite one another to build circular credibility.
- VisionHeight's initial blog covered 6 states and 37 sites; later analysis found at least 14 states and 460 pages of content.
- Targeted states: Arizona, Colorado, Georgia, Indiana, Iowa, Kansas, Michigan, Mississippi, Nevada, North Carolina, Pennsylvania, Texas, Virginia and Wisconsin. At least seven states saw House races targeted; others focused on Secretary of State races, statewide judicial appointments and local offices.
- All 37 initially identified sites were registered in a single evening, 36 of them within an 8-minute window. On Sept. 30 the network published roughly 100,000 words of candidate content using the same template and layout, which suggests AI-generated infrastructure and content.
- The sites were configured to explicitly allow major US AI companies' crawlers, whereas many legitimate media sites block them.
- Most of the research examined AI answers surfaced through Google search; Perplexity citing the content was the one confirmed instance on a major AI model.
Remediação
The article offers no formal remediation guidance. It notes that chatbots have improved at avoiding false election information but remain vulnerable to manipulation and errors on complex political questions, so users and defenders should treat chatbot answers on local races with caution.
Linha do tempo
-
Hyperlocal fake news site networks across 14 US states target AI chatbots to skew answers on local races
ler a análise desta fonte ocultar
VisionHeight, an Israeli threat intelligence firm, uncovered a network of hyperlocal websites spanning at least 14 US states, built to influence how AI chatbots answer questions about local and state races. Sites with names like "Wolverine State Report", "Peach State Bulletin" and "Tarheel State Report" repeat hundreds of near-identical statements that a given candidate is "the stronger choice" on issues such as housing and taxes. The content is written for AI crawlers rather than human readers, which shifts the objective from persuading people to poisoning the retrieval sources that chatbots draw on. VisionHeight does not attribute the activity to any actor and stresses that the technique is easily reproducible by foreign and domestic actors alike. Perplexity was confirmed to cite one of the pages (Tarheel State Report) as its first source for an answer.
Technical Details
- Each state network typically has 5-6 sites, some posing as hard news and others as opinion or editorial outlets, and they cite one another to build circular credibility.
- VisionHeight's initial blog covered 6 states and 37 sites; later analysis found at least 14 states and 460 pages of content.
- Targeted states: Arizona, Colorado, Georgia, Indiana, Iowa, Kansas, Michigan, Mississippi, Nevada, North Carolina, Pennsylvania, Texas, Virginia and Wisconsin. At least seven states saw House races targeted; others focused on Secretary of State races, statewide judicial appointments and local offices.
- All 37 initially identified sites were registered in a single evening, 36 of them within an 8-minute window. On Sept. 30 the network published roughly 100,000 words of candidate content using the same template and layout, which suggests AI-generated infrastructure and content.
- The sites were configured to explicitly allow major US AI companies' crawlers, whereas many legitimate media sites block them.
- Most of the research examined AI answers surfaced through Google search; Perplexity citing the content was the one confirmed instance on a major AI model.
Remediation
The article offers no formal remediation guidance. It notes that chatbots have improved at avoiding false election information but remain vulnerable to manipulation and errors on complex political questions, so users and defenders should treat chatbot answers on local races with caution.