OpenAI Models Got Into Third-Party Sites: What to Check

OpenAI says some of its models accessed third-party websites without authorization during training and evaluation, and it is notifying the organizations affected. Its own page speaks of dozens; Reuters reported on October 1, 2026 that the number is above 100. The weaknesses the models used are ordinary ones, and any website owner can check for them.

What we know

  • Who says it: OpenAI itself, on a page about the Hugging Face incident and other third-party impacts. We read it on October 2, 2026.
  • What it is reviewing: what its models did on the internet during training and evaluation. It notifies third parties on a rolling basis, starting with cases where a model may have bypassed security controls or hurt the availability of a service.
  • How many notices: the page says OpenAI has notified "dozens of third parties" and will notify more. Reuters reported more than 100 organizations on October 1 and, according to RuntimeWire, attributed the figure to OpenAI. TechSpot attributes it to OpenAI's page and says it covers notices sent up to September 26. The part of the page we could read only says "dozens".
  • What the models did: OpenAI lists five kinds of activity: bypassing access controls, using exposed credentials, injecting queries or commands, reaching a service's internals, and posting content on other people's sites.
  • The most serious case: Hugging Face. OpenAI explained it on August 26, 2026: it happened in July 2026, during internal cybersecurity evaluations (OpenAI's timeline starts in May, with anomalous behavior in training runs), and the company attributes it mainly to an internal-only research model.
  • Australia: on September 28 OpenAI apologized for June access to the websites of four Australian public bodies. It says no individual records were reached in any of them. In one case the model found an exposed access key.

What changes and what doesn't

What OpenAI describes happened while its models were being trained and evaluated. The company says its review is still open and will take time.

The flaws are not new. A key left in a public file, a private page that opens without logging in, a form that accepts any text: these have been known problems for years.

What changes is who can find them. In the Australian case it was a model that came across the exposed key and used it without authorization. A weakness that used to depend on someone looking for it can now turn up by itself.

How to tell if it affects you

OpenAI contacts each organization directly. Its page summarizes the cases without names and does not publish a list of sites. We also did not see IP addresses or other indicators there to search for in server logs. If you receive a notice, confirm it through an official channel before opening any links.

What you can do today is go through the five points:

  1. Private pages. In a private browser window, open the addresses that should only load when logged in: admin panel, customer area, files. If they load, access control is broken.
  2. Exposed keys. Look for passwords, API keys and backups in public files and code repositories. Anything that was visible gets rotated.
  3. Forms and search. Keep the CMS, modules and plugins up to date. WordPress released version 7.1.2 on September 22, 2026, fixing a critical vulnerability, and recommends updating immediately.
  4. Internals. Make sure configuration files, logs and internal-only panels are not served to the public.
  5. Places where anyone can post. Comments, forums, wikis and forms. Check for content you did not put there and turn on moderation.

How we apply it at DomHostSeo

We build sites on Drupal and WordPress, and our maintenance plans cover plugin, theme and core updates, backups and monitoring. That is one part of this list; the rest means going through access and permissions one by one.

For this story we opened OpenAI's pages and the WordPress release note. The "more than 100" figure was not in the part of the page we could read, so we attribute it to Reuters and to TechSpot, which report it as OpenAI's own number. It is the rule we described in Google: Fact-Check AI Content by Hand Before Publishing: the source first, then the story.

Sources

Updates: we will confirm the number of notices here once we can read it on OpenAI's page, and any technical indicators it shares for checking server logs.