GHSA-6688-9rhm-gjv2 DOMPurify: IN_PLACE returns a force-removed rawtext root whose text carries attacker markup — pure HTML reparse executes

• npm dompurify: < 3.4.16 → fixed in 3.4.16

Environment • dompurify 3.4.15 (current npm release); reproduced independently on jsdom 30.0.1 and 29.1.1 (Node.js 20.x / 26.x) • Config: `DOMPurify.sanitize(node, { IN_PLACE: true })` on a Node input; `SAFE_FOR_XML` at its default (`true`) Summary The 3.4.9 fix for the IN_PLACE detached-root class added two protections on the IN_PLACE return path: a fail-closed `TypeError` in `_forceRemove` when a node selected for removal cannot be detached, and a `_neutralizeSubtree` pass (`dist/purify.js` line 1336) that strips non-allowlisted attributes from removed subtrees. Both miss the rawtext text-content form. When the force-removed root is a rawtext element (``), the payload lives in the node's *text*: the node detaches fine (the `TypeError` guard is not reached), `_neutralizeSubtree` strips nothing (there are no attributes), and the IN_PLACE exit returns the detached, never-sanitized `` whose text still carries live markup. Serializing that node and re-parsing it in plain HTML context materializes the payload — no foreign-content context required. The same Node input sanitized without `IN_PLACE` returns an empty result: the only difference is the IN_PLACE return path handing the killed node back. Steps to reproduce ```js const { JSDOM } = require('jsdom'); const createDOMPurify = require('dompurify'); // 3.4.15 const window = new JSDOM('').window; const DOMPurify = createDOMPurify(window); const styleRoot = window.document.createElement('style'); styleRoot.setAttribute('onclick', 'alert(1)'); // attribute payload styleRoot.textContent = ''; // text payload window.document.body.appendChild(styleRoot); const returned = DOMPurify.sanitize(styleRoot, { IN_PLACE: true }); console.log(returned === styleRoot); // true (same node) console.log(styleRoot.parentNode === null); // true (detached) console.log(styleRoot.outerHTML); // console.log(styleRoot.getAttribute('onclick')); // null (attribute neutralized) console.log(styleRoot.textContent); // '' (text survives) // plain HTML reparse (no foreign-content context involved): const probe = window.document.createElement('div'); probe.innerHTML = returned.outerHTML || styleRoot.outerHTML; console.log(probe.querySelectorAll('img').length); // 1 console.log(probe.querySelector('img').getAttribute('onerror')); // "1" ``` Observed on 3.4.15: one node, one call — the `onclick` attribute is neutralized while the text payload (``) survives verbatim; serializing and re-parsing the returned node in plain HTML context materializes the `img` with the live `onerror` handler. Contrast on the same Node input without `IN_PLACE`: `RETURN_DOM: true` → ``; `RETURN_DOM_FRAGMENT: true` → 0 children — the payload is fully sanitized away. The only difference is the IN_PLACE return path. Contrast on the removal trigger: `SAFE_FOR_XML: false` → the node is not removed (detached stays false); plain CSS text → not removed. The removal is gated by the mXSS text probes and happens *specifically because* the serialized node would re-open tags on reparse. Root cause `_isUnsafeNode` (`dist/purify.js` 3.4.15, lines 1700–1714) removes nodes whose literal text would re-open tags on reparse — shape (b) in the source comment is "text-only content that already carries the element's OWN end tag", detected by the `LITERAL_TEXT_CLOSE` probe (line 385) alongside the `ELEMENT_MARKUP_PROBE` (line 339) rules. `_forceRemove` (line 1122) records the node in `DOMPurify.removed` (`{element}`) and detaches it. The removal is intentional: the upstream comment states these shapes are removed because the literal serializer emits them verbatim for the HTML parser to re-open. The IN_PLACE exit then hands the force-removed root back to the caller — the very node whose removal `DOMPurify.removed` just recorded (verified: `DOMPurify.removed.some(e => e.element === root)` is `true` on the returned instance). The 3.4.9 `_neutralizeSubtree` pass (line 1336) addresses only the attribute form — its own docstring: "walks a removed subtree and strips every attribute" (purpose: cancel queued resource events). Rawtext text content is out of its scope, so the removal that was performed *specifically to prevent reparse* is undone by returning the node: you removed it to stop the reparse, then returned it. Differential (one node, one call, same removal path): the `onclick` attribute is neutralized by the existing pass while the text payload survives verbatim — the attribute axis is covered, the text axis is the gap. Impact Identical blast radius to the published IN_PLACE family: an application that sanitizes a Node in `IN_PLACE` mode and re-inserts (or serializes and then re-inserts) the result materializes attacker markup in plain HTML context: script execution in the page. Moving the returned node via `appendChild` alone is safe; the round trip through serialization is what fires the payload. No foreign-content context is required with the close-tag payload. Affected versions • Verified live: 3.4.15 (current). • Source-verified: the attribute-only `_neutralizeSubtree` and the IN_PLACE return path are present in 3.4.9–3.4.14; releases before 3.4.9 predate the fix entirely (unconditional return; individual pre-3.4.9 releases not dynamically tested). • Per cure53 advisory convention the affected range is reported as ` e.element === root)` — and fail closed: throw the same `TypeError` style used by the 3.4.9 detach guard ("a node selected for removal could not be safely returned; refusing to sanitize in place"), or return `null`. This is consistent with the existing fail-closed design and covers all present and future root-kill reasons in one check. Secondary (form-specific): extend `_neutralizeSubtree` to neutralize text content of rawtext descendants — the elements in `LITERAL_TEXT_ELEMENT_NAMES` (`style`, `script`, `xmp`, `iframe`, `noembed`, `noframes`, `plaintext`, `noscript`) — by rewriting `textContent` to a defanged form, matching the probe coverage of `_isUnsafeNode`/`LITERAL_TEXT_CLOSE`. A regression test asserting that a force-removed rawtext root comes back with no `/<[/\w!]/` match in `textContent` (and is not returned at all under the primary fix) prevents re-introduction. Prior art / differentiation • GHSA-r47g-fvhr-h676 (fixed 3.4.6): clobbered-form root removal — different trigger; this report's root is a normal allowlisted `style` element killed by the text probe. • GHSA-55q2-fjhq-7xh7 (low): IN_PLACE hook removal leaves a detached subtree executable — the attribute-form twin (hook-stripped subtree retains onload-class handlers). This report's rawtext text form is not covered by `_neutralizeSubtree`'s attribute stripping and is not that advisory. • GHSA-h8r8-wccr-v5f2 (medium): mXSS via re-contextualization in the standard (non-IN_PLACE) serialize path — different mechanism; IN_PLACE is not involved. • The 3.4.9 release notes credit @mozfreedyb for the IN_PLACE handling improvements that this residual escapes on the text axis. Applicability scope (stated up front) The payload materializes when the application serializes and re-parses the sanitizer output (`innerHTML` assignment, template rendering, markdown/HTML round trips) or otherwise consumes the returned node's markup. Moving the returned node via `appendChild` alone does not trigger it. Applications that pass live, connected attacker trees into `IN_PLACE` are explicitly warned against by upstream's own source comment; this report concerns the serialize-and-reinser…