Is mechanicalsoup safe?

GHSA-x456-3ccm-m6j4

A malicious web server can read arbitrary files on the client using a inside HTML form.

This affects the extremely common pattern of form submission:

b = mechanicalsoup.StatefulBrowser()
b.select_form(...)
b.submit_selected()

The problem is with the code in browser.Browser.getrequestkwargs:

    if tag.get("type", "").lower() == "file" and multipart:
        filepath = value
        if filepath != "" and isinstance(filepath, str):
            content = open(filepath, "rb")
        else:
            content = ""
        filename = os.path.basename(filepath)
        # If value is the empty string, we still pass it
        # for consistency with browsers (see
        # https://github.com/MechanicalSoup/MechanicalSoup/issues/250).
        files[name] = (filename, content)

The file path is taken from the bs4 tag "value" attribute. However, this path will default to whatever the server sends. So if a malicious web server were to send something like:

then upon .submitselected() the mechanicalsoup browser will happily send over the contents of your SSH private key.

import attr
import mechanicalsoup
import requests


class NevermindError(Exception):
    pass


@attr.s
class FakeSession:
    session = attr.ib()

    headers = property(lambda self: self.session.headers)

    def request(self, *args, **kwargs):
        print("requested", args, kwargs)
        raise NevermindError  # don't actually send request


def demonstrate(inputs=None):
    b = mechanicalsoup.StatefulBrowser(FakeSession(requests.Session()))
    b.open_fake_page("""\







""", url="http://127.0.0.1:9/")
    b.select_form()
    if inputs is not None:
        b.form.set_input(inputs)
    try:
        b.submit_selected()
    except NevermindError:
        pass

# %%

# unpatched
demonstrate()
# OUTPUT: requested () {'method': 'post', 'url': 'http://127.0.0.1:9/', 'files': {'evil': ('passwd', ), 'second': ('', '')}, 'headers': {'Referer': 'http://127.0.0.1:9/'}, 'data': [('greeting', 'hello')]}

# %%

# with the patch, this now works. users MUST open the file manually and
# use browser.set_input() using the file object.
demonstrate({"greeting": "hiya", "evil": open("/etc/hostname", "rb").name, "second": open("/dev/null", "rb")})
# OUTPUT: requested () {'method': 'post', 'url': 'http://127.0.0.1:9/', 'files': {'evil': ('hostname', ), 'second': ('null', )}, 'headers': {'Referer': 'http://127.0.0.1:9/'}, 'data': [('greeting', 'hiya')]}

# %%

# with the patch, this raises a ValueError with a helpful string
demonstrate({"evil": "/etc/hostname"})

# %%

# with the patch, we silently send no file if a malicious server tries the attack:
demonstrate()

diff --git a/mechanicalsoup/browser.py b/mechanicalsoup/browser.py
index 285f8bb..68bc65e 100644
--- a/mechanicalsoup/browser.py
+++ b/mechanicalsoup/browser.py
@@ -1,7 +1,8 @@
+import io
 import os
 import tempfile
 import urllib
 import weakref
 import webbrowser
 
 import bs4
@@ -227,15 +228,21 @@ class Browser:
                     value = tag.get("value", "")
 
                 # If the enctype is not multipart, the filename is put in
                 # the form as a text input and the file is not sent.
                 if tag.get("type", "").lower() == "file" and multipart:
                     filepath = value
                     if filepath != "" and isinstance(filepath, str):
-                        content = open(filepath, "rb")
+                        content = getattr(tag, "_mechanicalsoup_file", None)
+                        if content is False:
+                            raise ValueError(
+                                """From v1.3.0 onwards, you must pass an open file object directly, for example using `form.set_input({"name": open("/path/to/filename", "rb")})`. This change is to mitigate a security vulnerability where a malicious web server could read arbitrary files from the client."""
+                            )
+                        elif not isinstance(content, io.IOBase):
+                            content = ""
                     else:
                         content = ""
                     filename = os.path.basename(filepath)
                     # If value is the empty string, we still pass it
                     # for consistency with browsers (see
                     # https://github.com/MechanicalSoup/MechanicalSoup/issues/250).
                     files[name] = (filename, content)
diff --git a/mechanicalsoup/form.py b/mechanicalsoup/form.py
index a67195c..82f6015 100644
--- a/mechanicalsoup/form.py
+++ b/mechanicalsoup/form.py
@@ -1,8 +1,9 @@
 import copy
+import io
 import warnings
 
 from bs4 import BeautifulSoup

All users of MechanicalSoup's form submission are affected, unless they took very specific (and manual) steps to reset HTML form field values.