Is ChatGPT agent safe? It's retired, and its replacement stays signed in as you
The Morse teamPublished 10 min read
As of November 2026, ChatGPT agent is retired, so the safety question now belongs to its replacement, ChatGPT Work's cloud browser. That browser keeps your password away from the model better than the old agent did, but it stays signed in to your accounts between tasks and keeps working after you close your laptop. OpenAI's help page now opens with "ChatGPT agent is no longer available," and OpenAI says the new browser's safeguards "do not eliminate every risk." It's reasonably safe for searching and comparing on public sites, and worth more care on any site where it's signed in as you.
We read OpenAI's ChatGPT agent help page, its cloud browser and ChatGPT Work help pages, the ChatGPT release notes, its data controls and chat deletion pages, and the ChatGPT agent system card on November 16, 2026, and used Internet Archive copies of the agent page to date the change. OpenAI deserves credit for publishing how often its agent resisted planted instructions in its own tests, including the cases where it didn't.
ChatGPT agent and its replacement: key facts, November 2026
| Setting | Default | Retention or limit | Source |
|---|---|---|---|
| ChatGPT agent (agent mode) | "No longer available" | Old chats and screenshots kept until you delete them, then removed within 90 days | OpenAI |
| Cloud browser in ChatGPT Work | Paid plans, "excluding Free and Go" | Runs on "a separate computer in the cloud" | OpenAI |
| Signing in to websites | Secure sign-in form; Plus and Pro | Session "will persist for future tasks until it expires" | OpenAI, release notes |
| Access to a new website | Asks you first | Can be changed to "Auto approve" or "Always allow" | OpenAI |
| Payments and bookings | Asks for confirmation in chat | Applies even on sites you allowed | OpenAI |
| Cookies and sign-ins | Kept in the cloud browser | Until you clear them under Settings, Cloud browser, Browser data | OpenAI |
| Training on Plus and Pro chats | On | Off under Improve the model for everyone | OpenAI |
| Chat history | Kept until you delete it | Permanent deletion scheduled "within 30 days" after you do | OpenAI |
Is ChatGPT agent still available?
No. The notice at the top of OpenAI's agent help page reads: "ChatGPT agent is no longer available. Use ChatGPT Work for longer, multi-step tasks and finished deliverables." For browser tasks it points to a page called "Using cloud browser in ChatGPT."
OpenAI didn't put a retirement date in its release notes, but the Internet Archive narrows it down. Its July 11, 2026 copy of the help page has no notice, and its July 20, 2026 copy has it. OpenAI's release notes introduced ChatGPT Work on July 9, 2026, as "an agent for longer, more involved tasks."
Below the notice, the old page is still there in full, with agent mode's monthly limits (40 messages on Plus, 400 on Pro) and the /agent command. Operator went the same way earlier. The agent page says "Operator functionality is now integrated into ChatGPT agent mode" and "the Operator website is no longer accessible." OpenAI retired its Atlas browser too, and its release notes say Atlas was "scheduled to stop working on August 9, 2026." We cover that in is ChatGPT Atlas safe.
Does ChatGPT agent see your passwords?
The old agent was built not to, and the cloud browser goes further. In agent mode, a login made the agent pause and ask you to choose "Take over browser." The agent help page says that "while you control the browser, screenshots are not captured," so it didn't record what you typed.
The cloud browser uses a secure sign-in form instead. OpenAI's cloud browser page says credentials entered there "go directly to the remote browser," that the username and password "are not visible to the model," and that "ChatGPT does not store those sign-in credentials." Before the form appears, "an additional review model" checks the request for signs of phishing, and you can see the site's address and a preview of the form. The August 25, 2026 release note that added sign-in says the browser "supports the use of password managers" and that your password is never "used in model training."
Typing a password into the chat itself skips all of this. The cloud browser page says "Never paste passwords, security codes, or payment details into the conversation," and the old agent page said to use takeover mode for anything sensitive.
Does ChatGPT's agent stay logged in to your accounts?
Yes. The old agent page says "cookies persist across sessions for convenience, just like a regular browser." The cloud browser page says that after you sign in, "the authentication will persist for future tasks until it expires."
The cloud browser runs on OpenAI's computers, not yours. OpenAI's cloud browser page says it keeps "its own cookies, browser data, and signed-in sessions," and that it doesn't touch your own browser's tabs, history, saved passwords, cookies, or extensions. That keeps your laptop's sessions out of it, but every site you sign in to through it leaves a live session on a machine you don't control.
To clear it, open Settings, select Cloud browser, then under Browser data select Cookies. You can clear every site or one at a time, and clearing a site signs ChatGPT out of it. If that session was your bank, sign it out first. Your bank's own terms decide who answers for what an agent does with your login, and our page on bank agreements covers what Chase and Schwab say.
Can ChatGPT's agent keep working when you're not watching?
Now it can, by design. When OpenAI launched the agent in July 2025, its system card described "Watch Mode," which on sensitive sites, "e.g. logged into an email or banking account," paused the agent "when the user becomes inactive or navigates away from the conversation."
The cloud browser is built to keep going. OpenAI's page says it "can continue in the background, including after you close your computer or turn off your phone," and pauses only "when it needs information, sign-in, or confirmation." The page doesn't mention watch mode.
The confirmation step remains. The cloud browser page says ChatGPT is "designed to request confirmation before actions that could be hard to reverse," such as "confirming a booking or making a payment," and allowing a site doesn't skip it. In OpenAI's own tests of the July 2025 agent, published in its system card, it asked first 100% of the time before completing a financial transaction and 99.9% of the time before sending high-stakes messages, but 91.0% across all actions that needed a confirmation. OpenAI notes that last figure "is an underestimate." Those tests are of the retired agent. OpenAI hasn't published the same numbers for Work.
How often did ChatGPT agent fall for prompt injection?
In OpenAI's own tests, it resisted most attacks but not all. Prompt injection is text on a page that gives the agent orders. The agent help page gives the example of a comment on a restaurant site telling the agent "to retrieve a password reset code from Gmail and send it to a malicious website." Our prompt injection explainer covers how these attacks work.
The system card's Table 12 gave these rates for the agent's visual browser, as of July 17, 2025:
| Attack in OpenAI's test set | ChatGPT agent resisted | Operator (o3) resisted |
|---|---|---|
| Instructions planted on a web page | 95% | 89% |
| Trying to leak data already in the chat | 78% | 80% |
| Trying to make the agent fetch data, then leak it | 67% | 75% |
The system card says these numbers test "only the model's behavior, not our full end-to-end stack," which also includes monitors, filters, and confirmations. At launch it also switched off ChatGPT memory for the agent, "to mitigate the risk of prompt injections attempting to exfiltrate data from memory." For the cloud browser, OpenAI's page says only that it "tests for risks such as prompt injection, phishing, and unintended actions," and that those safeguards "do not eliminate every risk."
How long does OpenAI keep what the agent saw?
Until you delete it. The agent help page says your agent chats, "browsing history in ChatGPT agent, and screenshots" stay in your history until you delete them, and that deleted chats and screenshots are removed "from our systems within 90 days." For ordinary chats, OpenAI's deletion page says it schedules permanent deletion "within 30 days."
People can read it. The agent page says agent content, "including screenshots," may be accessed by "authorized OpenAI personnel" and "trusted service providers" to investigate abuse, give support, handle legal matters, or improve models unless you've opted out.
For Work, the cloud browser page says information from websites "is handled according to your ChatGPT data-control settings." It doesn't give the cloud browser's screenshots a retention period of their own. OpenAI's Work page adds that "messages and task context may be stored in the cloud, even when work runs locally" on the desktop app.
Training follows your account. OpenAI's data controls FAQ says data sharing "is enabled by default" on Free, Plus, and Pro, and that Business, Enterprise, and Edu content isn't used for training by default. Our page on whether ChatGPT trains on your data has the steps to turn it off.
What changed from ChatGPT agent to the cloud browser?
Sign-in got safer, and supervision got looser.
| ChatGPT agent (July 2025 to July 2026) | ChatGPT Work cloud browser (November 2026) | |
|---|---|---|
| How you sign in | Take over the virtual browser and type | Secure sign-in form, password managers supported |
| Does the model see the password? | Screenshots paused while you type | "Not visible to the model," not stored |
| Phishing check before sign-in | Not described on its help page | A review model checks the request |
| Sessions | Cookies "persist across sessions" | Persist "until it expires" or you clear them |
| On email and banking sites | At launch, Watch Mode paused it if you left | Keeps running after you close your computer |
| New websites | Some sites blocked by OpenAI | Asks first, unless you choose "Auto approve" or "Always allow" |
| Plans | Plus, Pro, Business, Enterprise, and Edu | Paid plans except Free and Go; sign-in on Plus and Pro |
Sources: OpenAI's agent help page, system card, cloud browser page, and release notes, read November 16, 2026.
How do you use ChatGPT's cloud browser safely?
Give it less to work with, and keep it asking you first.
- Keep website access on "Always ask." OpenAI calls "Always allow" "not recommended."
- Give it one narrow task. The agent page warns against prompts like "Check my email and handle everything."
- Read each confirmation before you approve it. Check the site address and the amount, not only the summary.
- Clear the session afterward. After a task on your bank, insurer, or email, clear that site under Settings, Cloud browser, Browser data.
- Turn off apps the task doesn't need. Event-triggered Work tasks can run on new Gmail, Slack, or GitHub activity (OpenAI), so see how to disconnect apps from ChatGPT and what scheduled tasks can read.
- Delete the chat when you're done if it holds screenshots of an account page.
Some websites block automated browsers on purpose. OpenAI's cloud browser page says that's the site's decision, and suggests trying another website or opening the page in your own browser instead.
How to get help with a task without handing over a login
The cloud browser's risks come from an AI holding a signed-in session on someone else's computer, with screenshots of what it saw filed under your account. Most of what people hand an agent, like comparing utility plans, reading a bill, or drafting a reply to an insurer, can be done with the facts alone.
On Morse, you ask in words and paste only what the question needs. Your chat history is stored in your browser, encrypted, not on Morse's servers, and our servers don't keep your messages or log what you write. Morse doesn't train AI on your chats, sell your data, or show ads. Pick a mode for each chat:
- E2EE, for lines from the bill, statement, or email you'd otherwise let an agent open. Your browser checks the secure enclave's proof that it's genuine, then encrypts your message so only that enclave can read it, and Morse's servers relay ciphertext they can't read. The enclaves are run by third-party providers, not by Morse, and the protection rests on the enclave hardware and attestation working as designed. E2EE runs open-weight models with no web search or image generation, and comes with Pro at $19 a month. Here's how end-to-end encrypted AI chat works.
- Private, the default, for planning the task yourself. Work out the steps, compare the options you paste in, or write the form answers with open-weight models like DeepSeek, Kimi, and GLM, or with Grok. Morse has zero data retention agreements with its inference, search, and other providers, so they've agreed to keep nothing, and where a provider confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't.
- Anonymous, to have GPT, Claude, or Gemini draft the message without the provider knowing it's you. Our proxy removes your account, email, and IP address and sends the message under Morse's own account. The provider sees the text without knowing it's you and may keep it for a limited time to check for abuse, under its own policy, so leave account numbers and names out.
Clearing your browser, losing your device, or deleting your account deletes your chats, and Morse can't recover them, so copy out anything you'll need. The privacy page has the details.
Try your next errand as a question from the Morse home page without an account, or start free with no card.
Common questions
Is ChatGPT agent the same thing as Operator?
Operator came first. OpenAI's agent help page says Operator's features were folded into agent mode and the Operator website is no longer accessible, so both names now point to the same retired product.
Why won't ChatGPT's cloud browser open a website that works for me?
Some websites block automated browsers, and OpenAI says that choice is the website's, not ChatGPT's. OpenAI suggests trying another site or opening the page in your own browser.
Can I use ChatGPT Work's cloud browser on the Free plan?
No. OpenAI lists cloud browser for paid plans only, excluding Free and Go, and signing in to websites with it for Plus and Pro.
I used agent mode before it was retired. What should I delete?
Delete the agent chats that show account pages, since deleting a chat also deletes its screenshots and OpenAI then removes them within 90 days. If you also signed in to sites through ChatGPT Work, clear those sessions in its cloud browser settings.