Bento, the office suite that fits in a file
TypeScript mit-licenseofficeoffice-365office-suiteoffice-toolsoffice365open-sourcepowerpointself-hostedsingle-page-appslidesslideshareslideshowslideshow-galleryslideshow-makerslideshows
docs: the text sanitizer walks what it unwraps (#467) — DECISIONS entry and 1.1.0 changelog line (#472)
nyblnet95e6dadf831d
diff --git a/CHANGELOG.md b/CHANGELOG.mdindex 3e7d1841..548d3ce3 100644--- a/CHANGELOG.md+++ b/CHANGELOG.md@@ -13,6 +13,16 @@ pre-1.0.## [1.1.0] — 2026-09-14+- **Security: update this file. Text in a deck could run code when clicked+ or hovered.** Formatted text — a text box or a table cell — could carry+ script inside an ordinary-looking tag that the checker skipped over, and it+ ran when a reader clicked or moved the mouse across that text, in the editor+ and in the show. A deck or a pasted clip from someone else was enough;+ nothing looked wrong on screen. As with 1.0.16, anything running inside the+ page inherits what the page holds — the live-session keys, the local+ autosave copy, write access to the file where the browser grants it. Every+ shell before 1.1.0 is affected. Text is now checked at every nesting depth+ before it renders, and the check is proven by clicking, not by inspection.- **Live broadcast.** Contributed by Niemes (#293), and shaped together withthe collaboration work below. *Audience copy…* in the Share menu writes a hand-outfor a live show: whoever opens it lands straight in the presentation, anddiff --git a/docs/DECISIONS.md b/docs/DECISIONS.mdindex 76a95ed1..eb0d4d68 100644--- a/docs/DECISIONS.md+++ b/docs/DECISIONS.md@@ -6884,3 +6884,44 @@ toast. Measured: with the switch on, a click opens nothing.Also in the same change, unrelated to security: `* ` makes a bullet like `- `(#255), two or more leading spaces make an indented sub-bullet (#368), and the`?` overlay lists the shortcuts it had been missing (#269).++## 2026-09-14 — The text sanitizer walks what it unwraps++A hole in `sanitizeHtml` (`slides/src/render.ts`), the one sanitizer for+every text element and table cell — run at render, on the canvas, and on+every paste — was open in every shell since the function was written. Fixed+in #467 (`95b6b97`), shipped in 1.1.0. Written down because the shape is+general and a later sanitizer could repeat it.++**What it did.** The walk iterated a SNAPSHOT of a node's children. On a tag+not in the allowlist it unwrapped: lifted the children into the parent ahead+of the cursor, removed the tag, and `continue`d. The lifted children were not+in the snapshot, so the pass never visited them. Any allowed tag nested inside+any unknown tag kept every attribute it arrived with —+`<section><b onclick="…">`, `<foo><span style="position:fixed">`, at any+depth — while the same markup under an allowed parent was cleaned. Measured+in Chrome through the real text-element render path: a click and a hover on+such text ran the handlers. A deck file or a Bento clipboard payload from+someone else was enough; nothing looked wrong on screen.++**Why the rig did not see it.** Every case in `test-sanitize` was flat. The+sanitizer's own allowlist was right and every flat attack was refused, so the+rig was green for the whole of its life. The shape nobody had tried was+nesting an allowed tag inside a refused one.++**The fix.** Walk first, then lift: `walk(elChild)` before the children are+moved up, so a lifted child is held to the same rule as its siblings. Two+lines, one at the general unwrap and one at the refused-anchor unwrap that+#465 adds. `sanitizeSvg` was never affected — it REMOVES an unknown element+whole rather than unwrapping it, which is the other correct answer.++**The rig now.** One case renders five nested shapes plus the allowed-parent+control through `renderSlide`, then clicks and hovers the lifted elements in+Chrome. With the recursion removed it goes red three ways and the handlers+run (`pwned 41,42`); with it, nothing runs. That is the assertion that would+have caught this: not "did the attribute survive" but "did anything execute".++**The rule this leaves.** A sanitizer that removes an element has exactly two+correct options for its children: drop them with it, or sanitize them before+lifting them. Lifting first is always wrong, whatever the loop looks like.+And every attack case in a sanitizer rig gets a nested variant.