criticalaccess.aiAI for rural hospitals

Security and compliance audit · MIT license

Scanning hospital websites for hidden hacks.

A hospital website can be quietly hijacked with hidden gambling or spam links. An old hospital domain can be taken over. The hospital may never notice. Many sites also miss federal notice requirements. This series of audits looks at both, using only what any visitor can see.

Python 3.10+curldigInternet ArchiveBuilt with Claude Code

01 · At a glance

The question, the data, and the terms.

Three audits, run from July to November 2026, with tools anyone can reuse. No hospital is named on this page.

The question
Is a hospital’s website, or a domain it used to own, showing visitors something the hospital never put there? And does the site carry the notices federal rules call for?
Who it’s for
Hospital IT and compliance leaders, and researchers.
Data
Only what any visitor can see on public web pages, plus CMS roster data, the HIFLD hospital dataset, Internet Archive captures, and DNS and registration records. For the national run, hospital names and domains are left out of the repository.
Status
A completed audit series, July to November 2026, with reusable tools.
License
MIT.

02 · What it does

It reads the site the way a visitor does.

The scanner requests public pages and looks for the byte-level signatures of known injection campaigns: hidden links, cloaked content, and malicious JavaScript. The scanner itself makes no AI calls.

It also checks domains a hospital used to own. When a hospital changes its name or closes, the old address can lapse and be picked up by someone else, while old links and listings still point to it.

AI was used in two places around the scanner. One step used an AI model to judge archived copies of pages. The compliance reviews were done by AI agents against a fixed rubric.

Responsible useThe rules the audits followed

  • Passive requests only. Ordinary page requests, the kind any browser makes.
  • One request at a time per site.
  • Never bypass bot checks.
  • Report privately first. Hospitals are told before any public naming.

03 · What it found

Three audits, counted carefully.

The figures below are the project’s own, as published in its repository. In the national scan, every flagged site was checked by hand.

July to August 2026

Compliance audit

281 rural hospital websites in 9 states were checked against three rules: price transparency, the No Surprises Act, and ACA Section 1557.

  • 54 of the 281 (19%) met all three.
  • Price transparency was met by 211 (75%), the No Surprises Act by 115 (41%), and Section 1557 by 134 (48%).
  • 103 of 188 sites in the later runs had no Good Faith Estimate notice.
  • 134 of 281 lacked language-assistance taglines.
August to October 2026

National critical access hospital security scan

All 1,382 critical access hospitals in 45 states were covered. 1,075 distinct sites, and 44,524 pages, were scanned.

  • 16 sites were flagged and checked by hand.
  • 4 were confirmed compromised (0.37%).
  • 3 of the 4 were hospital sites with hidden gambling links. 2 were still live at the October recheck.
  • 1 was an abandoned hospital domain hosting gambling.
November 2026

Iowa statewide audit

123 hospitals were covered, with 85 live sites and 3,568 pages scanned. No live Iowa hospital website was serving injected content.

  • 22 were flagged and 2 confirmed. Both were former hospital domains now hosting gambling.
  • 51 of 58 former domains are still held by the hospital or its system.
  • A third former domain carried betting pages in 2021 and is now unregistered. A federal dataset still lists it as the hospital’s website.

04 · Limits

What “clean” does and doesn’t mean.

A passive scan sees only so much. These limits come from the project’s own documentation.

Limits
  • It samples pages: 40 per site, and up to 140 for large system sites.
  • It sees only what visitors see. A backdoor that does not change the rendered pages, or an outdated plugin, will not show up.
  • It detects only the injection campaigns the project describes. “Clean” means none of those was visible. It does not mean a site is uncompromised.
  • About one in ten sites was unreachable in the national run.
  • A wrong roster address produces convincing false findings, so each website has to be confirmed before anything is counted.
  • The compliance check reports where notices appear. It does not judge whether their wording satisfies each rule.

Check your own site and your old domains.

The tools are on GitHub under the MIT license. Follow the same rules the audits did: passive requests, one at a time, and tell the hospital privately first.