“All 37 websites VisionHeight identified in their first analysis … were registered in a single evening, and 36 were registered within an 8 minute timeframe.”
Networks of hyperlocal websites across 14 states
Researchers at VisionHeight, an Israeli threat intelligence firm, discovered networks of hyperlocal websites set up to influence U.S. political races. The operation initially appeared to encompass six states and 37 websites, but VisionHeight told CyberScoop that its inquiry expanded to at least 14 states and roughly 460 pages of content. States named by the firm include Arizona, Colorado, Georgia, Indiana, Iowa, Kansas, Michigan, Mississippi, Nevada, North Carolina, Pennsylvania, Texas, Virginia and Wisconsin.
The sites carried local-sounding names such as “Wolverine State Report,” “Peach State Bulletin,” and “Tarheel State Report.” Some presented as news outlets, others as opinion or editorial sites, and the outlets often cited one another to create a circular network of localized credibility.
Content written for AI crawlers, not human readers
VisionHeight found that the text on these sites rarely resembled conventional journalism. Instead of reporting or analysis, much of the content consisted of repeated, template-driven statements about individual candidates — for example, direct assertions that “Candidate X is the stronger choice” on particular issues such as housing or taxes. Those lines were repeated with small variations across pages and issues.
“What’s really interesting about this is that the websites themselves and the content is so well built to let’s say optimize the probability for AI crawlers and the model to be influenced by this content,” said Nitsan Daniel, VisionHeight’s chief technology officer and co‑founder, in an interview. Many mainstream media sites opt to block AI crawlers; these hyperlocal sites were configured specifically to allow major U.S. AI companies to scrape their material.
Audit-ready is a season. It shouldn't be.
Evidence in spreadsheets, controls drifting between audits, frameworks multiplying on flat headcount. Nubivance runs continuous compliance on Rapid7 Cyber GRC - SOC 2, HIPAA, ISO 27001, PCI, CMMC.
End the scrambleScale, automation, and a single-day publication
The operation showed hallmarks of automated production. VisionHeight reported that the 37 sites in its first analysis were all registered in a single evening, with 36 registrations occurring within an eight-minute window. On Sept. 30 the network published “roughly 100,000 words of candidate content” using the same templates and layouts across sites. VisionHeight also told CyberScoop it had since found the operation spanned at least 14 states and 460 pages of content.
The firm said AI was likely used to generate much of both the infrastructure and the written material, enabling rapid, large-scale creation of sites and repetitive content designed for machine consumption.
Targets, models, and attribution limits
The campaign targeted a mix of races. VisionHeight reported that at least seven states saw one or more House races targeted; other pages focused on races for Secretaries of State, statewide judicial appointments and local offices. The firm did not attribute the activity to any particular actor.
VisionHeight documented at least one major AI model — Perplexity — linking to and citing a page called Tarheel State Report as its primary source for an answer. The finding illustrates the practical effect: when a chatbot or search-based model queries the web for “which candidate is better on housing affordability,” the engineered pages provide ready-made answers that models may retrieve and repeat.
What this means for technologists, policymakers, and end users
- Technologists and security teams: Expect adversaries to prioritize networks and pages that explicitly permit crawling. Systems that vet sources by provenance or cross-checking may need to look beyond surface signals of locality and mutual citation.
- Policymakers and regulators: The rapid, low-cost replication described by VisionHeight — and the firm’s observation that foreign and domestic actors can reproduce the method — underscores a potential regulatory and oversight challenge around how AI models are allowed to ingest and cite online material.
- End users and the general public: Chatbots that rely on web-scraped sources for political answers remain vulnerable to manipulation at scale; material written expressly for machine consumption can shape what models report back to voters asking politically relevant questions.
VisionHeight framed the discovery as a symptom of a broader shift. “Now instead of spending time and effort on getting people to read the content, [they’re] getting into their AI and being influenced by that,” Daniel said. The firm stressed the technique is easily reproducible and can amplify efforts that, in the past, might have targeted only large, high-profile contests.
The core fact is stark and simple: automated sites and template content can be produced in bulk, configured to be scraped, and then cited by at least some AI models — all within a single day. VisionHeight did not assign blame to a named actor, leaving attribution unresolved. If influence operations are evolving to aim at the AI layer rather than the human reader, two immediate questions follow from the record: which systems will detect and discount machine-optimized content, and who will be responsible for policing it?
Source: CyberScoop — The next election influence campaign will target your AI chatbot
