ChatGPT’s Computer History has one big privacy problem

ChatGPT's Computer History feature on Mac allows the app to remember user activity across different applications and websites. However, OpenAI's documentation reveals a significant privacy issue - the files storing this history are not encrypted. Instead, interaction events are stored temporarily on the Mac and then converted into plain-text Markdown files. OpenAI specifically warns that these files are not encrypted and could potentially be accessed by other programs running under the same macOS user account.

This lack of encryption poses a considerable risk. While ChatGPT's Computer History does not record the screen, audio, or video, the interaction data can still reveal sensitive information such as which apps and websites a user interacted with and what documents they worked with. A malicious application running under the same user account could potentially access these files, giving it another source of information about the user's activity.

Moreover, there is a broader concern with providing AI systems with such extensive context. Malicious instructions hidden within websites or other content could potentially influence what an AI sees in its history, introducing a prompt-injection risk on top of the data exposure issue. Thankfully, Computer History is an opt-in feature, and OpenAI provides users with controls to exclude specific apps and websites and delete entries.

While this is an improvement over silently recording everything, the documentation highlights a trade-off: the more ChatGPT remembers about a user's Mac, the more sensitive information it potentially leaves behind. When that information is not encrypted, it creates a compromise worth considering before enabling the feature.