OpenAI has introduced a new feature for its ChatGPT desktop application on macOS, called Computer History, which is designed to create a timeline of a user's computer activity. This feature, which replaces an earlier research preview known as Chronicle, generates summaries of user interactions, noting the applications and websites involved.
Computer History is an opt-in feature, disabled by default, and requires OpenAI’s separate Memories feature to be active. It is currently available to Pro users, while Business and Enterprise users need administrator approval before they can individually opt in. The feature is not available in the European Economic Area, Switzerland, or the United Kingdom.
The system operates by processing interaction events and leveraging text and other contextual information accessible via macOS accessibility features. OpenAI states that it does not record screenshots or audio, and web browsing conducted in private mode is excluded from collection. Users can specify which applications and websites contribute to the history and can exclude others at any time. A pause function is available in settings and the macOS menu bar to temporarily halt data collection.
Users have control over their history, with options to clear the last ten minutes, hour, day, or the entire record. Each timeline entry's corresponding memory file can be located in Finder. OpenAI clarifies that raw interaction-event files remain on the user's Mac, isolated within the ChatGPT app's data container, and are deleted after 48 hours. The memory files generated from these events, which are plain-text Markdown documents, persist until manually deleted by the user. OpenAI also states that these event files are processed on its servers solely for summary generation and are not retained afterward.
OpenAI has identified two primary security risks associated with Computer History. First, the memory files are not encrypted, meaning other programs running under the same macOS user account could potentially access and read them. Second, feeding a large volume of running context to the model increases the risk of prompt injection attacks. This could occur if a user visits a website containing malicious instructions, which ChatGPT or Codex might then follow.
Cybersecurity experts have raised additional concerns. One analyst likened the feature to having a coworker meticulously log every click and keystroke, storing these notes unencrypted on the Mac. This raises questions about the appropriateness of enabling such a feature on work computers, particularly given the potential for sensitive information to be recorded. Another expert warned that these unencrypted memory files could provide a comprehensive map of a user's daily activities, making them a valuable target for information-stealing malware.
