ChatGPT Locked Chats Found in Code as Codex Brings Code Into the App

What the Android teardown found

OpenAI appears to be building a way to put individual ChatGPT conversations behind a PIN or fingerprint. The evidence comes from inactive code in version 1.2026.230 of the Android app, which Android Authority reported in late August6. The strings in that build are fairly complete. They describe a feature called "Locked chats" that keeps a conversation's title and content out of the sidebar, chat history and search results until the user authenticates6. Opening a locked chat does not unlock it for good. It stays locked until the user turns it back into a regular chat6.

A later Android beta added more detail. Locked conversations would sit under "Projects → Locked chats," and a new "Unlock chat" option would put a conversation back in the normal sidebar5. The wording in that version refers to a "PIN or device unlock method"5. Android Authority's onboarding strings say locked chats are "protected by an Account PIN"6. That small difference in wording points to a PIN tied to the ChatGPT account, not just to the phone's lock screen. If so, the lock could carry across devices. OpenAI has not confirmed that.

As of this writing, OpenAI has not announced the feature or given a timeline, and the code does not switch anything on for testers910. Users are still filing requests for this exact feature on OpenAI's developer forum, including one posted on October 927. That suggests it had not shipped broadly by early October.

Saved but hidden: a middle option next to Temporary Chat

The coverage agrees on how this differs from what ChatGPT offers now. Temporary Chats aren't saved to history, so you can't go back to them later6. Locked chats do the opposite: the conversation is kept, just out of sight38. Android Authority described this as fixing Temporary Chat's biggest weakness6. The draft onboarding text separates the two plainly: temporary chats aren't saved and don't create memories, while locked chats are saved and create memories kept apart from the user's general memory6.

That memory design is the most interesting part of the feature. According to the strings, locked chats can draw on what ChatGPT remembers from both regular and locked conversations. Anything new learned inside a locked chat can only shape other locked chats, never regular ones6. The code also warns that locking a chat doesn't erase what ChatGPT may already remember6. Crypto Briefing called this a separate memory sandbox for locked conversations3.

Outlets describe the memory behavior with different levels of certainty. Android Authority quoted the strings directly6. Others said it was unclear or only possible810. VARINDIA went further, saying information in a locked chat would stay inside that one conversation4, which is narrower than what the strings say. The direct quotes are the most reliable account: memory is shared among locked chats, not limited to one.

Some headlines also got ahead of the facts. Dataconomy said ChatGPT "introduces" locked chats7, and a syndicated version on a CoinDesk-branded site said the feature had been "added" to the Android beta2. Both articles go on to say nothing has launched. The CoinDesk-branded text admits the feature is still buried in code and not visible to beta testers27. The accurate summary is that locked chats are a feature in development, not a release.

Why developers should pay attention

The obvious users are people with private personal chats. Outlets mention health questions, money, relationships, and shared phones or tablets passed between family members7910. The case for developers is less obvious, and the timing makes it stronger.

ChatGPT now handles far more source code than it did a year ago. At DevDay on September 29, OpenAI added a Code Review plugin to the ChatGPT desktop app. It shows a pull request's description, changed files, comments and checks in one place, and it is available on all plans1820. Codex can also run automatic background reviews, so findings are ready when the developer sits down to review13. On the same day OpenAI launched a new Codex Cloud for Plus and higher plans, and Codex Security Cloud, which scans GitHub repositories, for Pro, Business, Enterprise and Edu customers20.

The result is that a developer's ChatGPT sidebar now holds conversations about unreleased features, security fixes and internal setup. Codex already has some safeguards in place. Cloud secrets are available to setup scripts but removed before the agent starts work11, and network access is off by default, both locally and in cloud containers15. Those safeguards limit what the agent can reach. They don't limit what a colleague sees when glancing at an open laptop, or what's visible on a shared device. A conversation-level lock fills that gap. A chat about a vulnerability that Codex Security flagged before the patch ships is a sensitive thread. A shared laptop at a reception desk or storefront is the kind of setting the feature appears designed for9.

What a lock does not protect

Locked chats should not be mistaken for data protection, and the better coverage makes that clear. Absolute Geeks pointed out that hiding a conversation in the interface is not the same as isolating its data, and that visual privacy, account privacy and data retention are separate things10. The CoinDesk-branded piece noted that OpenAI hasn't said whether locked chats get different treatment on its servers, such as exclusion from training or different retention rules2. One commenter on Tibor Blaho's post put it bluntly: the feature is closer to hiding than to privacy, and Temporary Chat is still the only mode that stores nothing5.

Even Temporary Chat has limits. A guide for lawyers from Spellbook notes that OpenAI may keep temporary conversations for up to 30 days for abuse monitoring. It adds that OpenAI holds the encryption keys, so stored chats can be read on OpenAI's side and handed over under legal compulsion26. Nothing found in the locked-chats code suggests that will change.

For code, the rules that matter are set at the account level, not by any interface feature. On individual ChatGPT plans, Codex content may be used for training unless the user opts out. Codex also has a separate setting for training on full environments, and changing the general ChatGPT privacy setting doesn't change it1112. On Business, Enterprise and the API, inputs and outputs aren't used for training by default, and zero data retention is available for eligible endpoints12. Aurascape argues that the real risk is developers putting company code into personal accounts12. A PIN on a sidebar entry does nothing about that.

The review workflow has its own exposure risks, separate from chat history. OpenAI's documentation warns that a comment posted in the Summary or Changes views of Code Review goes to GitHub immediately, without waiting for the reviewer to click Submit review17. Automatic reviews also need a ChatGPT plan; they don't work with API-key sign-in1819.

What it means

The best way to read locked chats is as a mature and somewhat overdue fix for screen-level privacy. It is not a security boundary. The feature follows WhatsApp's chat lock closely29. WhatsApp hides locked chats in a separate folder, hides names and message text in notifications, and offers a secret code for linked devices29. ChatGPT users have been asking for something similar since at least December 202424. The memory sandbox is the part that goes beyond a copy. It treats sensitive context as something to keep out of everyday personalization, not only out of view.

For developers, the takeaway is to use locked chats for discretion and to rely on account controls for security. If proprietary code is being discussed with Codex, the decisions that actually matter are made before the first prompt: use a workspace plan with training off by default, check the separate Codex training setting, and keep sensitive work local when the cloud setup doesn't fit1112.

Some questions are still open. No one knows yet whether a lock set on an Android phone will also hide the chat on the web or iOS, or what happens if a user forgets the PIN2. A lock that hides a chat on the phone but leaves it visible in a laptop browser wouldn't help much. The "Account PIN" wording points to sync across devices6, but that is a reading of draft text, not a confirmed design.