March 1, 2016
Django 1.9.3 fixes two security issues and several bugs in 1.9.2.
Django relies on user input in some cases (e.g.
django.contrib.auth.views.login() and i18n)
to redirect the user to an “on success” URL. The security check for these
redirects (namely django.utils.http.is_safe_url()) considered some URLs
with basic authentication credentials “safe” when they shouldn’t be.
For example, a URL like http://mysite.example.com\@attacker.com would be
considered safe if the request’s host is http://mysite.example.com, but
redirecting to this URL sends the user to attacker.com.
Also, if a developer relies on is_safe_url() to provide safe redirect
targets and puts such a URL into a link, they could suffer from an XSS attack.
In each major version of Django since 1.6, the default number of iterations for
the PBKDF2PasswordHasher and its subclasses has increased. This improves
the security of the password as the speed of hardware increases, however, it
also creates a timing difference between a login request for a user with a
password encoded in an older number of iterations and login request for a
nonexistent user (which runs the default hasher’s default number of iterations
since Django 1.6).
This only affects users who haven’t logged in since the iterations were increased. The first time a user logs in after an iterations increase, their password is updated with the new iterations and there is no longer a timing difference.
The new BasePasswordHasher.harden_runtime() method allows hashers to bridge
the runtime gap between the work factor (e.g. iterations) supplied in existing
encoded passwords and the default work factor of the hasher. This method
is implemented for PBKDF2PasswordHasher and BCryptPasswordHasher.
The number of rounds for the latter hasher hasn’t changed since Django 1.4, but
some projects may subclass it and increase the work factor as needed.
A warning will be emitted for any third-party password hashers that don’t
implement a harden_runtime() method.
If you have different password hashes in your database (such as SHA1 hashes from users who haven’t logged in since the default hasher switched to PBKDF2 in Django 1.4), the timing difference on a login request for these users may be even greater and this fix doesn’t remedy that difference (or any difference when changing hashers). You may be able to upgrade those hashes to prevent a timing attack for that case.
Skipped URL checks (new in 1.9) if the ROOT_URLCONF setting isn’t defined
(#26155).
Fixed a crash on PostgreSQL that prevented using TIME_ZONE=None and
USE_TZ=False (#26177).
Added system checks for query name clashes of hidden relationships (#26162).
Fixed a regression for cases where
ForeignObject.get_extra_descriptor_filter() returned a Q object
(#26153).
Fixed regression with an __in=qs lookup for a ForeignKey with
to_field set (#26196).
Made forms.FileField and utils.translation.lazy_number() picklable
(#26212).
Fixed RangeField and
ArrayField serialization with
None values (#26215).
Fixed a crash when filtering by a Decimal in RawQuery
(#26219).
Reallowed dashes in top-level domain names of URLs checked by
URLValidator to fix a regression in Django 1.8 (#26204).
Fixed some crashing deprecation shims in SimpleTemplateResponse that
regressed in Django 1.9 (#26253).
Fixed BoundField to reallow slices of subwidgets (#26267).
Changed the admin’s “permission denied” message in the login template to use
get_username instead of username to support custom user models
(#26231).
Fixed a crash when passing a nonexistent template name to the cached template
loader’s load_template() method (#26280).
Prevented ContentTypeManager instances from sharing their cache
(#26286).
Reverted a change in Django 1.9.2 (#25858) that prevented relative
lazy relationships defined on abstract models to be resolved according to
their concrete model’s app_label (#26186).
Dec 25, 2023