GHSA-rxcw-mc6f-6hr3

Описание

Summary

jodit's built-in clean-html sanitizer can be bypassed by a MathML/<style> carrier that hides a dangerous element from the sanitizer's element walk, so a no-interaction event handler survives into the editor value. When an application supplies attacker-influenced HTML to the editor's value-set or insertion paths, the sanitized output still contains a live <img ... onload=...> (or another non-onerror handler such as onfocus). A consumer that renders that output (element.innerHTML = editor.value) executes the handler with no user interaction. This is a stored cross-site scripting vulnerability, confirmed live on the shipped es2021/jodit.min.js for 4.12.25 and the latest 4.12.27 (in Chromium via a client-side innerHTML consumer, and in Firefox via server-rendered / document-context output; see the cross-browser note under Proof of concept).

Details

The bypass exploits the order in which clean-html parses, walks, and re-serializes the value.

1. On the value-set path, the clean-html plugin handles :beforeSetNativeEditorValue (src/plugins/clean-html/clean-html.ts:116), parsing the value into an inert document: sandBox.innerHTML = data.value.

2. In that parse, the source nesting math > mtext > table > mglyph > style triggers MathML text-integration-point and foster-parenting rules: the <img> is parsed as text inside <style> (rawtext), not as an element. The <table> is foster-parented out, and the <mglyph> MathML text-integration point governs the namespace, so the <img> never becomes an element node in this parse.

3. That value-set sanitizer is safeHTML (src/core/helpers/html/safe-html.ts:24); it walks the tree but acts on elements only (the Dom.isElement gate at :39) and runs against the parse-1 sandBox, in which the <img> is rawtext, not an element. So removeAllEventAttributes (the full on* strip at safe-html.ts:76) has no element to clean and the handler passes through. The onBeforeSetNativeEditorValue handler runs safeHTML on that parse-1 tree both before and after it captures the value, so neither pass ever sees the <img> as an element.

4. The captured value (data.value = sandBox.innerHTML) is then assigned to the editable, a second parse, which hoists the <img> out of <style> and into the HTML namespace as a live element with its handler intact. The serialize-reparse moves the element across the tree and across namespaces:

5. editor.value now carries that live element. The value-set pass (Steps 1-4) only walked the parse-1 sandBox and never sees it; the other sanitizer, the on-change visitor (visitNodeWalker via a LazyWalker, clean-html.ts:56/:70), does reach the hoisted element, but its sanitizeAttributes filter calls sanitizeHTMLElement (safe-html.ts:139), which strips onerror only - it never reads the removeEventAttributes flag sanitizeAttributes passes it (sanitize-attributes.ts:30), so it never runs the full on* strip. So onload, onfocus, and every other non-onerror handler is never removed and persists in editor.value permanently. (onerror is the one handler the cleaner removes, but only after a ~300ms window in which it too fires.)

The two code points (jodit 4.12.27):

All four layers are required: removing any of math + the integration point, table, the mglyph slot, or the rawtext element makes jodit strip the handler (substitutes per slot are under Carrier variants).

The four-layer carrier is required only on 4.11.2 and later. jodit 4.11.2 added cleanHTML.removeEventAttributes (the value-set full on* strip); before it (all 3.x and 4.0 through 4.10.x) the sanitizer only ever removed onerror, so on those versions a plain non-onerror handler such as <img ... onload=...> survives editor.value directly with no carrier (live-confirmed on 3.24.9, 4.0.1, 4.2.27). On 4.11.2 and later, the value-set walk strips the bare handler, so the carrier is needed to hide it as <style> rawtext past that walk; and because the on-change cleaner removes only onerror (Step 5), a non-onerror hoisted handler survives across the whole range.

The bypass is not specific to the value setter. The same carrier survives clean-html through editor.value = X, editor.setEditorValue(X), and editor.s.insertHTML(X) (the API jodit's own documentation uses for plugins and custom buttons).

This is distinct from jodit's known XSS advisories: CVE-2023-42399 (GHSA-95xr-cq6h-vwr3) is an iframe[src] URL-scheme issue fixed in a 4.0.0 beta, and CVE-2022-23461 (GHSA-42hx-vrxx-5r6v) is a paste-from-Word onerror desanitization in <= 3.24.2. Neither involves this parse-1 rawtext-hoist mechanism, and a search of the issue tracker for mglyph / mathml / mutation / "value xss" finds no prior report.

Proof of concept

Default configuration. Assign the payload, read it back, render it the way a consumer would:

Cross-browser: editor.value carries the hoisted live <img> in both Blink and Gecko, but the consumer's parse mode decides execution. Chromium fires it under both client-side element.innerHTML and document parsing; Firefox fires it only under document parsing (server-side rendering, document.write, <iframe srcdoc>), because under innerHTML Gecko leaves the <img> in the MathML namespace, inert. No iframe is required: a plain innerHTML consumer suffices in Chromium, any server-rendered consumer in Firefox.

Positive control, same run: a plain <img src=x onerror=alert(1)>, a plain <img onfocus=alert(1) autofocus tabindex=1>, and a plain <svg onload=alert(1)> are all stripped by the synchronous value-set pass to harmless output, proving clean-html is active; the same handlers pass through only when carrier-hidden as <style> rawtext, so the contrast isolates the rawtext-hiding step. (A plain <script> is a separate case - the value-set pass does not remove tags; only the async on-change cleaner removes it ~300ms later - so it is not part of this synchronous control.)

Carrier variants: three of the four layers accept substitutes: the MathML text-integration point (mtext / mi / ms / mn / mo), the integration-point child (mglyph / malignmark), and the rawtext element (style / xmp / noembed / script / plaintext; title / textarea / noscript do not work). math and table have no working substitute. The hidden element is not limited to <img>, and any non-onerror handler persists permanently. Element-name blocklisting will not close this.

Impact

Stored XSS with no user interaction, in the default configuration, on the input paths that jodit's own clean-html is responsible for sanitizing. jodit's own test suite asserts this is a sanitization boundary: src/plugins/clean-html/clean-html.test.js asserts that editor.value = '<p>test <img src="" onerror="alert(111)" alt=""></p>' sanitizes to <p>test <img src="" alt=""></p> (the onerror removed) under the default config. The carrier in this report passes that same default-config sanitizer yet keeps the handler live, defeating the asserted guarantee. A prior fix in 4.12.21 addressed a stored XSS premised on an application re-rendering editor.value as trusted HTML, so this threat model is maintainer-acknowledged. Precondition: an attacker can place HTML into the editor (a content-submission role) and the editor output is later rendered. Any integration binding value to application state in jodit-react, loading a previously stored document, or inserting content via a plugin/button will execute attacker script in the victim's page.

Suggested fix

Remove the gadget element at the source rather than re-sanitizing the output. A re-sanitize loop does not close this: nesting the carrier inside itself surfaces one level per parse, so one extra pass is bypassed at depth 2, and any fixed cap N is out-nested at depth N+2 (depth generalizes trivially). The robust fix is to drop any HTML-namespace element smuggled inside <math>/<svg> outside a spec integration point (<foreignObject>, <annotation-xml>, <desc>, <title>) during the element walk, the same approach DOMPurify's _checkValidNamespace uses. This is one pass and depth-independent. It must run on both entry points: the synchronous value-set safeHTML pass and the on-change walker (which does not call safeHTML). A regression test asserting that no on* survives a round-trip through editor.value, including the nested-carrier case, locks it in.