Today, revocation means hunting through each app: disconnect the integration, delete the memory, revoke the OAuth grant in your Google or Microsoft account page, and hope the copies die too. It works, slowly, one product at a time, and you can never confirm any of it.
Three places to check for any AI product: its own settings page for memory and personalization, your connected-apps pages (Google, Microsoft, GitHub) for OAuth grants it holds, and its data-deletion request flow for everything else. Do that for every assistant and agent you have ever tried. Most people stop at the first one, which is why old grants pile up for years.
A chatbot you stop using mostly just goes quiet. An agent keeps acting: reading mail, drafting, booking, spending. Authority you granted once persists until you find the right settings page. As agents multiply, "I revoked it everywhere, I think" becomes a genuinely dangerous sentence.
This is the revoke button Pupul exists for. Your context lives in one file you own, AIs read it by grant, and a single tap kills every grant at once. Any AI can verify the revocation offline against a signed status list, so a token that still looks valid is provably dead. There is a sixty-second demo where an AI loses access mid conversation. That gap between valid signature and dead access is the whole idea.
Take the free four-minute read, get your word and your one-of-a-kind eye, and add your pupil to ChatGPT or Claude in about two minutes. Revoke it all with one tap, anytime.
Take your read, free Watch the revoke demoNo. Revocation stops future access. Whatever the app copied while it had access stays unless you separately request deletion, which is why granting a readable file beats letting apps build their own copies.
Honest AIs check the status list and stop, and the file itself tells any reader to honor revocation. What revocation provably ends is verified access; it also ends the live verify link that made the file trustworthy.