OpenAI warns ChatGPT’s Computer History could create new risks

OpenAI has added a new ChatGPT feature that can remember parts of what a person does on a Mac. Called Computer History, it creates a timeline from activity across selected apps and websites so ChatGPT and Codex can understand recent work.

The idea could make an AI assistant more useful, but it also gives the system access to a wider slice of a person’s computer activity. OpenAI itself warns that the feature can capture sensitive information and increase the risk of prompt injection, which is why its privacy settings and limits matter.

What Computer History actually records

Computer History does not take screenshots or make screen recordings. Instead, it creates an interaction-event stream from apps and websites that a user allows. Those events can include clicks, typing, keyboard shortcuts, app switches, and other contexts available through macOS accessibility features.

The feature periodically turns that activity into text summaries and local memory files. ChatGPT and Codex can then use those memories to answer questions about recent work, help someone return to an unfinished task, or identify a document they were using earlier.

Private-mode browser activity is excluded, and the feature does not record microphone input or system audio. That makes it different from systems built around frequent screenshots. Still, recording typing and other computer interactions can reveal meaningful context, which is why OpenAI advises users to choose included sources carefully.

Why OpenAI says it could be useful

The main goal is to reduce the amount of background information people need to provide to ChatGPT. A user could ask what they were working on before a break, where a recent proposal is located, or which tasks they handled during the day, using the timeline as added context.

Computer History can also spot repeated patterns. If a person follows the same steps for a regular task, the system may suggest turning that workflow into a reusable skill or automation. The user still reviews the suggestion before asking Codex to create anything from the recorded workflow.

That memory can help ChatGPT work across multiple apps rather than relying only on the current chat. The trade-off is that better context requires broader access to activity data, so usefulness depends on how much information a person is comfortable with the system collecting.

The privacy concern is in a broader context

The risk is not limited to a single click or keystroke. A stream of small events can reveal what someone was reading, writing, or doing across the day. OpenAI warns that Computer History files may contain sensitive personal, financial, or health information if those sources are not excluded.

OpenAI also advises users to turn the feature off during communications with other people unless they have prior express consent. That warning matters because activity from messaging apps, work tools, or websites could include information about someone who never chose to share that material with an AI system.

The company provides users with controls to reduce that exposure, including app and website permissions, pausing, deleting, and full shutdown. The more sources a person includes, the more kinds of activity may appear in summaries or memories, so careful source selection remains important.

Prompt injection creates a second risk

OpenAI’s documentation identifies another problem, prompt injection. This happens when malicious instructions hidden in content try to influence an AI system. With Computer History, material in an allowed app or website can become part of the context later used by ChatGPT or Codex.

A person might visit a page that contains instructions intended for an AI rather than a human reader. OpenAI says Computer History increases this risk because the system may encounter such content while building context from computer activity, then potentially follow those instructions later.

That does not mean every webpage or document is dangerous. It means the feature expands the amount of outside content an AI may process as context. Users who enable it should be selective about sources and avoid assuming that activity captured from trusted-looking pages is automatically safe.

OpenAI built in several controls

Computer History is off by default. ChatGPT Pro users can choose to enable it, while Business and Enterprise customers need a workspace administrator to grant access before individual members can opt in. The feature also requires Memories to be turned on.

Users can choose only certain apps and websites, or exclude specific sources they do not want recorded. They can pause collection from the macOS menu bar, turn the feature off entirely, and inspect timeline items after they have been created.

History can be cleared on an individual item or across larger time periods, including the last 10 minutes, an hour, a day, or all history. OpenAI says clearing an item removes the related interaction events and memories. These controls make Computer History optional, but users still have to manage them.

What happens to the activity data

Computer History first stores interaction events temporarily on the Mac. OpenAI says the event files are isolated inside the ChatGPT App Group, which limits access from other apps without permission. ChatGPT and Codex delete those temporary event files after up to 48 hours.

The events are also processed on OpenAI’s servers to create memories. OpenAI says it does not retain temporary event files after processing unless required by law, and it does not use them for training. The resulting memory files are then stored locally on the Mac.

Those local memory files remain until the user deletes or clears them. OpenAI warns that Computer History does not encrypt these files itself and that other programs running as the same macOS user may access them. That makes device and account security important.

Availability is still limited

Computer History launched in the ChatGPT desktop app on macOS and is available to Pro, Business, and Enterprise customers. Business and Enterprise administrators must enable access before members can individually decide whether to turn the feature on.

The initial rollout does not include the European Economic Area, Switzerland, or the United Kingdom. OpenAI also says Computer History is not available through an API key or Amazon Bedrock, so the feature currently applies to a specific set of desktop ChatGPT users.

For those users, the choice comes down to convenience versus additional data exposure. Computer History can make ChatGPT more aware of unfinished work and repeated tasks. Still, OpenAI’s own warnings show why careful permissions, strong device security, and selective use are important before turning it on.

TL;DR:

• ChatGPT’s new Computer History feature can build a timeline from allowed Mac activity, including clicks, typing, keyboard shortcuts, and app switching, so ChatGPT and Codex can understand recent work.
• The feature is off by default, does not record screenshots or audio, excludes private browsing, and gives users controls to pause, limit, inspect, or delete the activity collected.
• OpenAI warns that Computer History may capture sensitive information and advises turning it off during communications with other people unless they have given prior express consent.
• OpenAI also says that computer history increases the risk of prompt injection because malicious instructions in apps or websites could be incorporated into the context used by ChatGPT or Codex.
• Computer History is currently available on macOS for ChatGPT Pro, Business, and Enterprise users, with administrator approval required in Business and Enterprise workspaces.

Don’t forget to follow us for more exclusive content.

If you liked this, you might also like:

• ChatGPT integration at Starbucks sparks user concern
• Study finds ChatGPT reliance may come with memory drawbacks
• OpenAI introduces task scheduling and web monitoring inside ChatGPT

This article was made with AI assistance and human editing.