On Windows, if an application is registered as an URL handler like this:
HKEY_CLASSES_ROOT
https
URL Protocol = ""
[...]
shell
open
command
(Default) = ".../qutebrowser.exe" "%1"
one would think that Windows takes care of making sure URLs can't inject
arguments by containing a quote. However, this is not the case, as
stated by the Microsoft docs:
https://docs.microsoft.com/en-us/previous-versions/windows/internet-explorer/ie-developer/platform-apis/aa767914(v=vs.85)
Security Warning: Applications that handle URI schemes must consider how to
respond to malicious data. Because handler applications can receive data
from untrusted sources, the URI and other parameter values passed to the
application may contain malicious data that attempts to exploit the handling
application.
and
As noted above, the string that is passed to a pluggable protocol handler
might be broken across multiple parameters. Malicious parties could use
additional quote or backslash characters to pass additional command line
parameters. For this reason, pluggable protocol handlers should assume that
any parameters on the command line could come from malicious parties, and
carefully validate them. Applications that could initiate dangerous actions
based on external data must first confirm those actions with the user. In
addition, handling applications should be tested with URIs that are overly
long or contain unexpected (or undesirable) character sequences.
Indeed it's trivial to pass a command to qutebrowser this way - given how
trivial the exploit is to recreate given the information above, here's a PoC:
https:x" ":spawn calc
(or qutebrowserurl: instead of https: if qutebrowser isn't registered as a
default browser)
Some applications do escape the quote characters before calling
qutebrowser - but others, like Outlook Desktop or .url files, do not.
As a fix, we add an --untrusted-args flag and some early validation of the raw
sys.argv, before parsing any arguments or e.g. creating a QApplication (which
might already allow injecting Qt flags there).
We assume that there's no way for an attacker to inject flags *before* the %1
placeholder in the registry, and add --untrusted-args as the last argument of
the registry entry. This way, it'd still be possible for users to customize
their invocation flags without having to remove --untrusted-args.
After --untrusted-args, however, we have some rather strict checks:
- There should be zero or one arguments, but not two (or more)
- Any argument may not start with - (flag) or : (qutebrowser command)
We also add the --untrusted-args flag to the Linux .desktop file, though it
should not be needed there, as the specification there is sane:
https://specifications.freedesktop.org/desktop-entry-spec/desktop-entry-spec-latest.html#exec-variables
Implementations must take care not to expand field codes into multiple
arguments unless explicitly instructed by this specification. This means
that name fields, filenames and other replacements that can contain spaces
must be passed as a single argument to the executable program after
expansion.
There is no comparable mechanism on macOS, which opens the application without
arguments and then sends an "open" event to it:
https://doc.qt.io/qt-5/qfileopenevent.html
This issue was introduced in qutebrowser v1.7.0 which started registering it as
URL handler: baee288890 / #4086
This is by no means an issue isolated to qutebrowser. Many other projects have
had similar trouble with Windows' rather unexpected behavior:
Electron / Exodus Bitcoin wallet:
- http://web.archive.org/web/20190702112128/https://medium.com/0xcc/electrons-bug-shellexecute-to-blame-cacb433d0d62
- https://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2018-1000006
- https://medium.com/hackernoon/exploiting-electron-rce-in-exodus-wallet-d9e6db13c374
IE/Firefox:
- https://bugzilla.mozilla.org/show_bug.cgi?id=384384
- https://bugzilla.mozilla.org/show_bug.cgi?id=1572838
Others:
- http://web.archive.org/web/20210930203632/https://www.vdoo.com/blog/exploiting-custom-protocol-handlers-in-windows
- https://parsiya.net/blog/2021-03-17-attack-surface-analysis-part-2-custom-protocol-handlers/
- etc. etc.
See CVE-2021-41146 / GHSA-vw27-fwjf-5qxm:
https://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2021-41146https://github.com/qutebrowser/qutebrowser/security/advisories/GHSA-vw27-fwjf-5qxm
Thanks to Ping Fan (Zetta) Ke of Valkyrie-X Security Research Group
(VXRL/@vxresearch) for finding and responsibly disclosing this issue.
(cherry picked from commit 8f46ba3f6d)
# Conflicts:
# qutebrowser/qutebrowser.py
# tests/unit/test_qutebrowser.py
If visiting an URL and then later using :back, we accidentally opened the last
URL in the back/forward history rather than the current one.
See #5359Fixes#5566
(cherry picked from commit 8f073ee095)
If lineno is set to None in the LogRecord, pytest's logging formatting fails:
tests/unit/utils/test_log.py:418: in test_empty_message
log.qt_message_handler(QtCore.QtDebugMsg, self.Context(), "")
qutebrowser/utils/log.py:508: in qt_message_handler
qt.handle(record)
/usr/lib/python3.8/logging/__init__.py:1587: in handle
self.callHandlers(record)
/usr/lib/python3.8/logging/__init__.py:1649: in callHandlers
hdlr.handle(record)
/usr/lib/python3.8/logging/__init__.py:950: in handle
self.emit(record)
../../pytest/src/_pytest/logging.py:326: in emit
super().emit(record)
/usr/lib/python3.8/logging/__init__.py:1089: in emit
self.handleError(record)
/usr/lib/python3.8/logging/__init__.py:1081: in emit
msg = self.format(record)
/usr/lib/python3.8/logging/__init__.py:925: in format
return fmt.format(record)
../../pytest/src/_pytest/logging.py:89: in format
return super().format(record)
/usr/lib/python3.8/logging/__init__.py:667: in format
s = self.formatMessage(record)
/usr/lib/python3.8/logging/__init__.py:636: in formatMessage
return self._style.format(record)
../../pytest/src/_pytest/logging.py:185: in format
return self._fmt % record.__dict__
E TypeError: %d format: a number is required, not NoneType
According to typeshed, lineno should never be None:
028f0d5293/stdlib/2and3/logging/__init__.pyi (L386)
Thus, this is our fault, not pytest's. However, before pytest 6.0.0, pytest did
not surface logging errors:
https://github.com/pytest-dev/pytest/pull/7231b13fcb23d7
Thus, we never noticed something was going wrong here.
(cherry picked from commit e206d346ce)
Make mypy happy
It doesn't know about QMessageLogContext.lineno being Optional[int] rather than
int.
(cherry picked from commit 4136c847fd)
For some reason, the event filter stopped inhibiting the context menu with Qt
5.9 (or possibly 5.8).
(cherry picked from commit 0ed215d81f)
Remove unnecessary return
(cherry picked from commit 8159206576)
This allows running mkvenv.py without having Pygments installed system-wide.
Prepending to the PATH (rather than appending) so that the virtualenv's one is
always used, so the system-wide one can be older or broken.
(cherry picked from commit 362d9d917b)
For some reason, __file__ was "scripts/dev/../asciidoc2html.py", causing
.parents[1] to point to "scripts/dev/" instead of "scripts/".
See #5534
(cherry picked from commit eed0373feb)
After 86d7943203, this now also re-adds the
workaround for QTBUG-58982 (#2290) because it is apparently still needed with
Qt 5.7...
This reverts commit 87d7dd9342.
How the handle width of 5 is distributed on macOS seems like totally random.
Instead of trying to get our calculations to work, just force the handle width
to an even number.
This was removed in f5f3bf63b5 but is still
needed so that the underlying QWebEngineView is actually deleted with Qt <
5.12. Otherwise, test_spawning_an_editor_and_closing_the_tab in
test_editor_bdd.py would fail because the element doesn't actually vanish.
Florian Bruhin
Wayne Cheng
Julin S