What you want to do
Trust that Poki stops and asks before a payment, a deletion, a sent message or a call, in any app and any language Poki ships in, and that an automatic approval is only ever given for a step I asked for myself.
What happens today
Two parts of the approval path were read in the Nov 14 code audit of main (0f89475) and 0.4.0 Beta 1. Nothing here was reproduced on a phone.
Sensitivity rules (native/ios/PokiCore/Sources/PokiCore/Approvals.swift, the word lists and matches())
- A Latin word counts only when it opens the label or the label has four words or fewer. A label such as "Complete purchase of 3 items" is treated as an ordinary tap.
- The lists have English and Simplified Chinese only. The app ships Traditional Chinese, and labels such as 購買, 刪除, 訂閱, 轉賬 are not matched. Japanese, Korean and Spanish labels are not covered either.
- Sending words are checked in ten listed messaging apps only.
- The guarded app list holds Wallet alone. Banking and brokerage apps get no extra care.
- Common action words are absent, for example confirm, submit, order, book, donate, withdraw.
- A button with no label never matches.
Auto-approval (native/ios-app/Poki/Sessions/ApprovalClassifier.swift)
- The route is chosen by whether the owner's chat or voice is active, not by who asked for the step. A step from an outside agent that arrives while the owner is chatting is judged with the owner's conversation as context.
- The prompt includes the approval summary and app name, which come from the screen. Text on a page or in a message is therefore read by the model that decides.
- The answer is accepted on an
ALLOWprefix. - Scope today is limited to sending, calls and password typing. Payments and deletions always ask, and a phone that never chose asks for everything, which is the right default.
Area
Poki iPhone app or Poki Setup
Proposal
Either direction helps, and they combine.
Widen the rules:
- Match a keyword anywhere in a short label, and at a word boundary in longer ones when the element is a button.
- Add Traditional Chinese now, then the other languages, and the missing action words.
- Treat sending words as sensitive in any app when the element is a button next to a text field, not only in the listed apps.
- Guard finance apps by App Store category or a maintained list, and let the owner add apps.
- Treat an unlabeled button on a screen that shows an amount or a secure field as sensitive.
Harden auto-approval:
- Let the model judge only steps that come from the owner's own chat or voice session. Outside agents always get the confirmation card.
- Mark screen-derived text in the prompt as untrusted data, and keep it short.
- Require the answer to be exactly one word.
- Record in the Activity Log which steps were approved automatically and why.
Related: the batched approval wording from the #269 review, where one Allow covers every sensitive step in a batch while the prompt names only the first.
Activity
4Thanks for confirming this from the code. Yes, please turn those examples into failing cases in the suite that covers matches() ("Complete purchase of 3 items", "Tap here to complete the payment", a weak word masking a stronger one later in the label, and a hyphenated label read as short). A PR with only the failing tests is welcome. We plan the approval fix for 0.4.2, together with granting only the first sensitive step of a batch.
Thanks @ZhehanZhang, @mgkvibez and @JosssphZhou for the concrete examples and continued discussion. I opened #372 for a confirmed part of this issue: long action-button labels and Traditional Chinese keywords could miss the sensitive-action gate. The regression also checks that a matching tap is not sent before approval. All 91 focused approval, sensitivity and fault tests pass locally, and all four PR checks pass. No physical-device or live purchase test was performed.
This is a partial fix, so I am leaving this issue open. Request provenance for auto-approval, single-step approval scope, unlabelled controls and wider financial-action coverage still need follow-up. If you have button labels or sanitized examples that still slip through, please share them; they would help extend the regression cases. Thanks again for helping make this boundary clearer.
Thanks for the careful audit, and to everyone who added cases.
Status in two parts. #372 (long action labels and Traditional Chinese keywords) is approved on review. It is now being checked here, with the PokiCore tests and a payment prompt on a real phone, and it will be merged after that passes. #380 holds the failing label cases, and it can be rebased once #372 lands.
This issue stays open for the rest: where an automatic approval request comes from, granting only the first sensitive step of a batch, unlabeled controls, wider coverage of finance apps and other languages, and tightening the auto-approval prompt. Those are planned.
Confirmed the matching half of this from the code on main, and it is exactly as you describe.
matches()inApprovals.swiftonly accepts a Latin word when the boundaries hold andopens || short, whereopensmeans every character before the match is not a letter or a number, andshortmeans the label is four whitespace-separated tokens or fewer. So "Purchase 3 items" hits, and "Complete purchase of 3 items" does not, even though "purchase" sits on clean word boundaries. The same rule explains why "Tap here to complete the payment" slips through while a bare "Pay" does not.Two consequences worth having in the issue while it is open:
reason()returns on the first matching word it finds, so a label containing a weak word ("open") can mask a stronger one later in the same string, and the token count uses whitespace splitting, so a hyphenated label counts as one token and reads as short.I have not reproduced anything on a device, same as you. If you want, I can turn the two examples above into failing cases in whatever suite covers
matches()today, since that part is pure string logic and does not need a phone.