Interaction logging arrives on macOS
OpenAI rolled out an update for its native macOS ChatGPT application, introducing an opt-in feature named Computer History. Designed for Pro, Business, and Enterprise subscribers, the system acts as a background context engine that translates daily user activity into structural memories readable by ChatGPT and Codex.
Rather than taking periodic desktop screenshots, the tool listens directly to operating system interaction events. It tracks window switches, keystroke frequency, shortcut usage, and file navigations to construct a unified chronological index of what the user is working on.
Key architectural differences from Windows Recall
The announcement immediately drew comparisons to Microsoft's Windows Recall tool. However, OpenAI implemented distinct design parameters to address long-standing surveillance concerns:
- Opt-in by default: Computer History remains strictly disabled until explicitly activated by the account owner inside the desktop settings.
- Zero visual media capture: The software refrains from saving screen captures or recording microphone audio, recording purely textual event data instead.
- Local timeline storage: Raw logs stay stored on the Mac file system, though interaction data is transmitted to cloud servers to generate structured memory embeddings.
- No model training: OpenAI stated that logs processed for memory generation are not retained on remote servers or utilized to train future models.
During internal testing, early users noted that ChatGPT could answer queries like «Find the document draft I was editing yesterday before my video meeting» by analyzing past event patterns.
The unencrypted risk hiding behind local context
While avoiding screen recordings reduces visual exposure, security researchers highlight a critical vulnerability embedded within the feature documentation. OpenAI confirmed that Computer History files stored locally on macOS are currently unencrypted, leaving the raw activity log readable by other third-party applications running on the same system.
This design creates a target for automated prompt injection. If a user unknowingly opens a web page containing hidden instructions, those malicious prompt strings can enter the Computer History timeline and silently manipulate subsequent AI task recommendations without triggering operating system alerts.