What happened
OpenAI added a feature called Computer History to the ChatGPT desktop app for macOS. According to OpenAI’s own documentation, it “turns your activity across apps and websites into memories and a timeline that ChatGPT and Codex can reference.” Codex is OpenAI’s coding agent, the tool that reads, writes, and runs code on your behalf when you ask it to build or fix something.
Here’s how it actually works. Computer History builds what OpenAI calls an “interaction-event stream” from apps and websites you’ve allowed it to watch. That means clicks, typing, keyboard shortcuts, switching between apps, and other activity that macOS exposes through its built-in accessibility system, the same system screen readers and automation tools use to see what’s on your screen. OpenAI’s documentation is specific that it does not capture screenshots, screen recordings, microphone input, or system audio. It’s a log of actions, not a recording.
That log turns into two things. First, a searchable timeline you can review yourself. Second, context that Codex can pull from directly. If you were editing a file, testing a script, or reading documentation an hour ago, Codex can find that activity and read it to pick up where you left off, instead of you re-explaining what you were doing. OpenAI also says ChatGPT can suggest a reusable skill or automation when it notices you doing the same kind of task repeatedly.
Computer History is off by default. Pro subscribers can turn it on themselves. Business and Enterprise accounts need an administrator to grant access before anyone on the team can enable it. It only works inside the ChatGPT desktop app itself, not for developers building on OpenAI’s API or using OpenAI’s models through Amazon’s cloud platform, and it isn’t available yet in the EEA, Switzerland, or the UK. Private browsing activity is never included, matching what OpenAI’s Ari Weinstein said publicly when the feature launched.
You get real controls over it: pause collection from the menu bar at any time, choose which specific apps or sites are included, or exclude ones you don’t want tracked. You can delete individual timeline entries, or clear the last 10 minutes, hour, day, or everything at once. OpenAI’s documentation notes that deletion is irreversible once you do it.
Why it matters
The most important line in OpenAI’s own documentation isn’t a feature description, it’s a warning: enabling Computer History “increases the risk of prompt injection from content in apps and websites.” Prompt injection is when hidden text in something an AI reads, a webpage, a document, a chat message, contains instructions the AI follows as if you had typed them yourself, instead of treating them as content to summarize or ignore.
That risk is real here because Codex isn’t just reading a summary you wrote. It’s reading raw activity your Computer History logged from wherever you were working, including websites. If one of the sites or apps you’ve allowed contains manipulated or malicious content, and Codex pulls that activity in as context for a coding task, it’s possible for instructions buried in that content to influence what Codex does next, potentially including actions on your codebase.
This is the same category of risk BuilderWithin has covered with Claude Code, where the permission prompts that ask you to approve a command before it runs exist specifically because an agent acting on unreviewed input can do real damage. Computer History adds a new pathway for unreviewed content, whatever you clicked on or read, to reach an agent that can execute code.
Who should care
Anyone using Codex on macOS through a ChatGPT Pro, Business, or Enterprise plan, especially if you’re the one deciding whether to turn Computer History on for yourself or your team. If you already juggle several apps and tools while coding and want Codex to carry context between them without you re-explaining it each time, this feature does that. It also widens what content can reach Codex without you reviewing it first.
What builders should do next
Before turning Computer History on, decide the exact list of apps and sites you’ll allow it to watch first, not after. OpenAI’s own controls let you include only specific apps and sites rather than everything by default; start narrow, with your editor, terminal, and maybe one or two trusted reference sites, rather than opening it to your whole browser history. Every site you add to that list is a new source Codex can end up reading from.
Check the timeline itself periodically, not just once. Open Computer History’s review view and look for entries from sites or apps you don’t recognize as sources you deliberately included. If something unexpected shows up, that’s your signal to remove it from your allowed list, since deleting a timeline entry afterward won’t undo any context Codex already pulled from it.
Run one real comparison before you rely on it daily. Pick a bounded task you’d normally hand to Codex, like resuming a specific code change or bug fix you stepped away from. Do it twice on the same day: once in a fresh Codex session with no Computer History context, and once with Computer History enabled for that session. Compare two things: how many follow-up corrections you had to make to get Codex oriented on the right file and goal, and whether the Computer History run pulled in any unrelated site or app content you didn’t expect to see referenced. If it only saves you a sentence of re-explaining but pulls in content you didn’t intend to share, the trade may not be worth it for that workflow.
End of article