lxml_html_clean CVE-2026-49825

lxml_html_clean.Cleaner does not strip javascript: URLs from namespaced URL attributes (xlink:href)

Reporter: Guillem Lefait <guillem@datamq.com> · Date: 2026-05-10 Affected: lxml ≤ 6.1.0 and lxml_html_clean ≤ 0.4.4 (latest stable) Confirmed against: lxml 6.1.0 + lxml_html_clean 0.4.4 on Python 3.13.5, 3.14.4, and 3.15.0a8 (libxml2 2.14.6 / 2.9.14 - bug is in pure-Python sanitizer logic, independent of the libxml2 backend) Root-cause class: same as CVE-2021-28957 (formaction missing from link_attrs)

Summary

Cleaner filters URL schemes (javascript:, vbscript:, …) by walking links via rewrite_links(), which delegates to iterlinks(), which only yields attributes named in lxml.html.defs.link_attrs. That allow-list contains no prefixed names (xlink:href) and no srcset. As a result, when Cleaner is configured with safe_attrs_only=False - a documented option for callers that want lenient attribute handling but still expect URL-scheme scrubbing - <a xlink:href="javascript:…"> survives sanitization untouched, and any browser that follows the SVG-anchor specification will execute the JavaScript when the rendered link is clicked.

CWE: CWE-79 (XSS), with CWE-184 (Incomplete List of Disallowed Inputs) as the underlying defect class.

Affected components

The legacy lxml.html.clean module - bundled in lxml < 5.2.0 and still installable on newer versions via the lxml[html_clean] extra - shares the same bug.

Root cause

defs.link_attrs is a flat string set; the literal xlink:href is absent:

# lxml/html/defs.py
link_attrs = frozenset([
    'action', 'archive', 'background', 'cite', 'classid',
    'codebase', 'data', 'href', 'longdesc', 'profile', 'src',
    'usemap', 'dynsrc', 'lowsrc', 'formaction',
])

HtmlMixin.iterlinks() (lxml/html/__init__.py:526-528) only yields attributes whose key is in that set:

for attrib in link_attrs:
    if attrib in attribs:
        yield (el, attrib, attribs[attrib], 0)

Cleaner.__call__ registers the URL-scheme filter via rewrite_links (lxml_html_clean/clean.py:348), which is a thin wrapper around iterlinks(). Because xlink:href is never yielded, _remove_javascript_link (clean.py:576) is never invoked for it.

Minimal reproducer

from lxml import html
from lxml_html_clean import Cleaner

for payload in (
    '<svg><a xlink:href="javascript:alert(1)">x</a></svg>',
    '<math><a xlink:href="javascript:alert(2)">y</a></math>',
):
    tree = html.fromstring(payload)
    Cleaner(safe_attrs_only=False)(tree)
    print(html.tostring(tree).decode())
    print('  iterlinks:', list(html.fromstring(payload).iterlinks()))
# <svg><a xlink:href="javascript:alert(1)">x</a></svg>     ← unchanged
#   iterlinks: []                                          ← link rewriter blind
# <math><a xlink:href="javascript:alert(2)">y</a></math>   ← unchanged
#   iterlinks: []                                          ← link rewriter blind

Both SVG and MathML scopes are vulnerable - same allow-list gap, both render anchors that browsers treat as navigable. Other lab-confirmed surviving variants (same scope, different scheme encoding): mixed-case (JaVaScRiPt:), HTML-entity (javascript:), embedded tab (java\tscript:).

Impact

A caller that uses Cleaner to neutralise untrusted HTML and chooses safe_attrs_only=False - typically because the application wants to allow custom data-/aria-/vendor attributes - will silently pass javascript: payloads carried on xlink:href through to victim renders. Stored XSS in any application that round-trips user-supplied HTML through this configuration. Reach is conditional on the safe_attrs_only=False toggle, but that is a documented public option; consumers reasonably expect URL-scheme scrubbing to be independent of attribute allow-listing.

Suggested fix

Extend link_attrs to include xlink:href. In HTML mode, lxml.html keeps prefixed attribute names verbatim - the parsed key is the literal string xlink:href, not a Clark-notation form - so the existing allow-list lookup is a plain string match. Same shape as the CVE-2021-28957 fix:

# lxml/html/defs.py
 link_attrs = frozenset([
     'action', 'archive', 'background', 'cite', 'classid',
     'codebase', 'data', 'href', 'longdesc', 'profile', 'src',
     'usemap', 'dynsrc', 'lowsrc', 'formaction',
+    'xlink:href',
 ])

This single change closes the reported XSS for both SVG <a xlink:href> and MathML <a xlink:href>. lxml_html_clean is the canonical home of the Cleaner code (881 lines); lxml.html.clean is a 21-line backward-compat shim (from lxml_html_clean import *) that picks up the fix automatically once link_attrs is updated upstream. Since the upstream change requires lxml maintainer action, see the alternative below if a self-contained patch in lxml_html_clean is preferred.

Alternative (in-package fix, no lxml coordination needed): add a namespaced-URL-attribute walk inside Cleaner.__call__ so the URL-scheme filter doesn't depend on link_attrs. Sketch:

# lxml_html_clean/clean.py - supplements rewrite_links() in __call__
_NS_URL_ATTRS = ('xlink:href',)
# extend as needed
_BAD_SCHEME = re.compile(r'^\s*(javascript|vbscript|data):', re.I)

for el in doc.iter():
    for attr in _NS_URL_ATTRS:
        if attr in el.attrib and _BAD_SCHEME.match(el.attrib[attr]):
            del el.attrib[attr]

This decouples the cleaner from the upstream link_attrs set and matches the security-ownership boundary established when the cleaner was extracted in lxml 5.2.0.

Defense in depth (optional, regardless of which fix path is taken):

• Also handle srcset: the value is a url 1x, url 2x, … descriptor list, so split on commas and validate each candidate URL. Not directly executable in current browsers, but closes the same gap.
• Also accept Clark-notation forms ({http://www.w3.org/1999/xlink}href) so XML-mode callers using lxml.etree get the same protection. HTML mode never produces this form, so not needed for the reported bug.

Severity

CVSS 3.1 base score: 8.2 / High - AV:N/AC:L/PR:N/UI:R/S:C/C:H/I:L/A:N (stored XSS; victim must click the SVG anchor; scope-changed because script executes in the rendering origin). PR:N reflects the common case where untrusted HTML enters the sanitizer from anonymous sources (comments, support tickets); deployments that gate writes behind authentication can score with PR:L (→ 7.6).

Severity is CONDITIONAL on the caller passing safe_attrs_only=False. With the class default (True), attribute allow-listing strips xlink:href before scheme scrubbing runs, and the bug does not fire - verified at HEAD: default-config Cleaner()(<svg><a xlink:href="javascript:…">x</a></svg>) → <svg><a>x</a></svg>.

Prior art / novelty

• CVE-2021-28957 (lxml 4.6.3) - same root cause, different attribute (formaction). Fix was a one-line extension of link_attrs. Direct precedent.
• CVE-2022-34473 (Mozilla Sanitizer API) - xlink:href URL bypass primitive in a different sanitizer.
• Bleach (Mozilla, Python) explicitly handles the xlink namespace; enshrined/svg-sanitize (PHP) ships cleanXlinkHrefs(); DOMPurify scrubs xlink:href via ALLOWED_URI_REGEXP.
• nh3 (the alternative recommended in lxml_html_clean's own README for security-sensitive use) is not vulnerable to this primitive - verified 2026-05-10 on nh3==0.3.5: with <svg>/<math>/<a> and xlink:href explicitly added to tags/attributes, both SVG and MathML payloads, all four scheme-encoding variants, are stripped (output e.g. <svg><a rel="noopener noreferrer">x</a></svg>).

Coordination

Filing as a private GHSA at fedora-python/lxml_html_clean - lxml_html_clean is the canonical maintainer of the Cleaner code (881 lines) and the security-responsible team since the lxml 5.2.0 split, where the cleaner was extracted out of lxml precisely so cleaner-security reports could land on the right team. The lxml side cannot be filed via GHSA (https://github.com/lxml/lxml/security/advisories/new returns 404 - private reporting is not enabled), so a parallel report has been emailed directly to the lxml maintainer for the upstream defs.link_attrs patch path. You're welcome to coordinate with them directly if you'd prefer the upstream fix over the in-package alternative above.

Happy to provide a draft patch or PR on either path. No bounty expected.