- A speaker on OpenAI's DevDay session said Sites hosts 8 million sites, and that over 70% of OpenAI's employees make their own.
- Our reading: where anyone can prompt an app into existence, each app is software that needs permissions, updates and an owner.
A speaker on OpenAI's DevDay 2026 session, "Sites and the Next Generation of Web Apps", described how easily staff now build and share small web apps. The speaker said Sites, a hosting service launched this year, already hosts 8 million sites. The speaker did not say whose sites those are. The way it was introduced suggests the figure covers Sites overall, not only OpenAI's internal apps.
What was said
A speaker said the product began as a personal fix. An engineer found it frustrating to deploy internal tools for researchers. So they built an easier route straight from Codex, OpenAI's coding agent. Other engineers passed the prototype around. Then non-engineers adopted it, and apps spread through the company.
A speaker said over 70% of OpenAI employees now make their own sites. They replace slide decks and spreadsheets and build tools for each team. The speaker listed prototypes, production tools and games, then said "we share those without sweating the details." The excerpt does not make clear whether that word means games, prototypes or everything built.
The demo showed how little stands in the way. A spreadsheet became a launch tracker from one prompt. A speaker said no skills or knowledge are needed: you prompt it and get a link to share. Sites also use each viewer's existing connectors. A speaker added: "Don't have to think about permissions or access."
The demo also showed write access. A speaker updated a Linear ticket through their own connector. A speaker also explained that anyone in the workspace with editor access can push updates to the hosted site.
Why it matters
Our reading: each of those conveniences is also a place where a weakness can hide. A site can refresh itself on a schedule. Editors can push changes to it. Viewers open it with their own access.
That raises a question the excerpts do not answer: what can an edit do with a viewer's access? It is a question, not a finding.
Building takes one prompt, so the count of apps can outrun anyone's knowledge of what exists. Managers who buy or allow such tools should ask who owns each app. They should also ask who can change it, what it connects to, and what happens when its maker leaves.
The other side
The excerpts do not show OpenAI skipping review. The speakers did not describe a review process, but they did not rule one out. The session is also a product demo, so it stresses speed and ease. It gives no incident, no failure rate and no security data.
The design has a plausible defence. Reusing each user's connectors means a site fetches what that viewer may already see. The demo showed this with two people seeing different tickets. Whether that holds across millions of sites is not answered here.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- OpenAI: Sites and the Next Generation of Web Apps | DevDay 2026 (2026-11-11)





