OpenAI's Computer History turns your Mac activity into Codex memory
Opt-in on Mac, it logs clicks and keystrokes so ChatGPT and Codex recall your work — and adds a prompt-injection surface. Plus: a Claude watermark detector.
Copy markdown
Codex and ChatGPT now read your work history
Computer History — opt-in, and replacing the old Chronicle preview — builds a searchable timeline from your clicks, typing, shortcuts and app switches via macOS accessibility, with no screenshots. Codex and ChatGPT tap that timeline as context to answer “what was I doing before lunch?”, find a doc you had open, or suggest an automation.
Off by default, and your memories sit in plaintext on disk
You have to switch it on, and Business/Enterprise admins gate it before anyone can. Raw events buffer locally for up to 48 hours, get processed on OpenAI's servers into plain-text memory files, then stay on your Mac until you delete them. Private browsing, microphone and system audio are excluded.
The catch: a wider prompt-injection surface
OpenAI itself warns that replaying your activity feeds more untrusted text into the model — content you merely viewed can smuggle in instructions. It advises pausing capture during sensitive work and excluding apps that hold secrets or credentials.
Update: Claude's watermark finally gets a detector
This morning's Nowline noted Claude's new text watermark shipped with no way to verify it. On Aug 14 Anthropic announced a detection API for third parties, launching soon — but it stays weak on short text, code and heavily edited output, and can't separate a full Claude write from a light touch-up.