OpenAI's Computer History writes a plaintext log of the workday. Decide before staff opt in.

OpenAI has shipped Computer History, a feature in the macOS ChatGPT desktop app that converts recent computer activity into memories that ChatGPT and Codex can use ^[1][2]. The consequence for endpoint and identity owners is narrow and concrete: the feature produces unencrypted plain-text Markdown files describing what a person did, and those files stay on the laptop until someone deletes them ^[3][4]. The mechanics are unremarkable and that is the point. Computer History builds a timeline out of everyday use, grouping activity into summaries and noting which apps and websites fed each one ^[5]. OpenAI says it draws on interaction events plus text and other context exposed through macOS accessibility features, and that it replaces the earlier Chronicle research preview as a rebuilt system rather than a rename ^[6][7]. The company says it does not include screenshots or record audio, and that private-mode browsing is never included ^[8]. The gating is where administrators should pay attention. The feature is off by default and opt-in, and it requires the separate Memories feature to be on ^[9]. Pro users can switch it on themselves; Business and Enterprise users need administrator approval before they can opt in individually, and approval alone does not enable it for anyone ^[10]. In practice the admin action is a one-time permission to let individuals decide, which means the policy question has to be settled before the first approval request, not after users start clicking yes ^[2]. It is also unavailable in the European Economic Area, Switzerland, and the United Kingdom, so geography answers the question for some tenants and not for their US staff ^[11]. The retention design is worth reading twice. OpenAI says raw interaction-event files stay on the Mac, isolated in the ChatGPT app's data container, and are deleted after 48 hours, and that the same files are processed on its servers only to build summaries and not kept afterward ^[12][13]. The derived memory files persist indefinitely ^[3]. So the artifact with the sandbox and the timer is the short-lived one, and the artifact that survives is the unencrypted, human-readable summary ^[1]. OpenAI flags two risks itself. The memory files are not encrypted, so other programs running under the same macOS user account may be able to read them ^[4]. And the running context raises prompt-injection exposure: "if you visit a website containing malicious instructions, ChatGPT or Codex might follow those instructions," the company warned ^[14]. Mark Beare, head of Malwarebytes' consumer security unit and Malwarebytes Labs, warns the files could give infostealers a ready-made map of someone's workday ^[15]. Read together, the stated protection boundary is the local user account, which is the boundary commodity stealers are built to operate inside ^[3]. Ed Gaile, Principal Solution Architect at Appfire, put the test plainly to Help Net Security: "Would you let a coworker sit next to you all day, write down every click and keystroke, and keep the notes in a folder on your Mac?" ^[16]. His threshold was documentary: "If you would have to defend that log in a room, a client memo, a personnel note, a spreadsheet with real numbers, I would not turn it on for the work computer" ^[17]. His advice was to leave it off the work Mac and use a personal machine with nothing that would need explaining ^[18]. There are real controls if you allow it. Users pick which apps and websites contribute and can exclude any of them, a pause button in settings and the macOS menu bar halts collection without disabling the feature, and history can be cleared by the last ten minutes, hour, day, or entirely ^[19]. Each timeline entry's memory file can be revealed in Finder ^[20], which is also the fastest way to hand your DLP and EDR teams a path to watch.