ChatGPT Work agents add signed-in website sessions in cloud browser

• ChatGPT Work can now reuse authenticated web sessions: OpenAIDevs says users take over the cloud browser to log in, then hand control back to the agent, with the login persisting across sessions.
• The change widens cloud browser beyond the launch limitation, because OpenAI's cloud browser help page still says it only works on public pages and cannot sign in at launch, while OpenAIDevs now describes signed-in website use.
• Work is the non-code spread of the Codex harness: OpenAI's launch post says it uses Codex technology, while pvncher described it as Codex on a cloud computer that can run code, debug, and publish Sites.
• The identity surface got bigger fast: Harmonic's security guide says signed-in browser state, screen contents, and clipboard can become agent-visible, and NickADobos claimed Sites can proxy out of the Work network sandbox.

OpenAI's cloud browser doc still carries the older no-sign-in rule; OpenAIDevs says the agent can now keep an authenticated session. The desktop browser doc already had a separate sign-in path with browser state, autofill, password management, extensions, downloads, and navigation. The run picker in _samirism's screenshot exposes the local/cloud split, and dkundel's mobile example shows why the cloud version matters: a Drive download became a voting Site from an Uber.

OpenAI framed the update as a human takeover followed by agent continuation. The user signs in inside the cloud browser, then the agent resumes the task with a session that persists across future runs, according to OpenAIDevs.

That changes the job class for cloud Work. OpenAI's cloud browser documentation says the launch version stops when a site requires credentials, autofill, password managers, sign-in, or payment; the new dev post says authenticated sites are now in scope.

The same doc says cloud browser runs remotely, can continue in the background, and pauses when it needs more information or confirmation. It also says cloud browser is available to paid ChatGPT plans except Free and Go in supported regions, with availability varying during rollout.

The UI now exposes two execution targets: local computer and cloud. thsottiaux said Work was available globally for paid plans across mobile, web, and desktop, while OpenAI's desktop app migration doc says cloud Work chats sync across web, mobile, and desktop.

OpenAI's built-in browser docs draw the sharper boundary:

• Cloud browser: remote browser for delegated web tasks that can continue in the background.
• Desktop built-in browser: browser inside the ChatGPT desktop app, visible to the user, with sign-in, autofill, password management, extensions, downloads, and navigation.
• Codex Chrome extension: path for existing Chrome profile, cookies, signed-in sessions, open tabs, or Chrome extensions.

The desktop app now packages Chat and Work under ChatGPT, alongside Codex on macOS and Windows, according to OpenAI's migration doc. Codex itself is not selectable on web or mobile, but supported desktop Codex chats are accessible from the Remote tab in the ChatGPT mobile app.

OpenAI's launch post says ChatGPT Work can gather context from apps and files, stay with a project for hours, and create sheets, slides, docs, and web apps. The same post says more than 5 million people use Codex weekly, with more than 1 million using it for work outside software development.

TheRealAdamG's inventory described the Work runtime as a cloud computer with 15 GB RAM, persistent files, terminal and code execution, remote browser, connected plugins, scheduled tasks, GitHub work, deploys, docs, sheets, slides, inbox and calendar summarization, and website monitoring.

A concrete mobile run came from dkundel, who said Work downloaded photos from Google Drive, then switched mid-task into a ChatGPT Site gallery with a voting system and used browser use to verify it. In a follow-up, dkundel said cloud execution fit tasks that do not need the user's own computer, such as daily task sync and document generation.

OpenAI's Sites documentation says Sites can be created in Work on ChatGPT web, and in Work or Codex inside the desktop app. Sites can be previewed, published, and shared as interactive websites or lightweight apps.

The signed-in cloud change arrives after desktop users were already seeing browser-state plumbing. moritzkremb's screenshot showed an import flow for Chrome passwords and cookies, and MakeUseOf's hands-on test said the built-in browser offered to import passwords and cookies from Chrome before the writer skipped it.

Harmonic's security guide frames browser state as an agent identity surface. It says signed-in browser state, screen contents, and clipboard can become agent-visible; it also names plugins, connectors, browser use, computer use, and scheduled tasks as separate governance surfaces.

OpenAI's desktop browser docs use similar language around untrusted web content. They say ChatGPT asks before using a new website unless the host was already allowed, and that cookies, logged-in status, and browsing history can be cleared from desktop app settings.

NickADobos called one Sites workaround a security issue and said patching it is difficult because handing information between a Site and the chat is part of the product idea.

The most interesting community claim was about Sites, not login. NickADobos said ChatGPT Sites can act as a network proxy to escape the ChatGPT Work network sandbox.

OpenAI's Sites docs say Sites are generated from Work or Codex, then previewed, published, and shared. That makes Sites both an output surface and an execution-adjacent surface when a Work agent builds a web app during a task.

NickADobos tied the claimed proxy behavior to that handoff surface in a follow-up, arguing that the hard part is preserving the core Sites behavior while closing the boundary gap.