[Python-Dev] PEP 466 (round 3): Python 2 network security enhancements

[Nick Coghlan]

And time for round 3 :)

Notable changes:

- removed the higher level networking modules from the scope of the
PEP. They made people nervous, and aren't actually needed to achieved
the desired result (at the higher levels, it's much easier for third
party pure Python modules like requests to step in and fill the gap,
as long as the core SSL infrastructure has been updated)
- dropped the sha module from the exemption and explicitly noted that
both the sha and md5 modules have been deprecated since Python 2.5
- conditional exemptions are now only granted for things that the
security infrastructure depends on, rather than the other way around
- answered the OpenSSL question based on MAL and Ned's feedback
- added the "2.7-legacy-ssl" branch to the proposal (along with some
pointed comments about the development priorities implied by
preferring this branch to the one with the improved security
infrastructure)
- made it even clearer that I consider the situation commercial
redistributor's responsibility to clean up, based on the support
commitments they have made to their users. The PEP is primarily about
changing one specific aspect of our maintenance policies so that the
onus is clearly pushed back on them to fill the gap between what their
customers want and what upstream volunteers are prepared to do for
free - we shouldn't give them the opportunity to hide behind the
excuse that we won't let them fix it upstream.

Cheers,
Nick.

====================================
PEP: 466
Title: Network Security Enhancement Exception for Python 2.7
Version: $Revision$
Last-Modified: $Date$
Author: Nick Coghlan <ncoghlan at gmail.com>,
Status: Draft
Type: Informational
Content-Type: text/x-rst
Created: 23-Mar-2014
Post-History: 23-Mar-2014


Abstract
========

Most CPython tracker issues are classified as errors in behaviour or
proposed enhancements. Most patches to fix behavioural errors are
applied to all active maintenance branches.  Enhancement patches are
restricted to the default branch that becomes the next Python version.

This cadence works reasonably well during Python's normal 18-24 month
feature release cycle, which is still applicable to the Python 3 series.
However, the age of the standard library in Python 2 has now reached a point
where it is sufficiently far behind the state of the art in network security
protocols for it to be causing real problems in commercial use cases
where upgrading to Python 3 in the near term may not be practical.

Accordingly, this PEP relaxes the normal restrictions by allowing
enhancements to be applied in Python 2.7 maintenance releases for standard
library components that have implications for the overall security of the
internet. In particular, the exception will apply to:

* the ``ssl`` module
* the ``hashlib`` module
* the ``hmac`` module
* the components of the ``random`` and ``os`` modules that the above
  modules rely on for cryptographic randomness
* the version of OpenSSL bundled with the binary installers for Windows
  and Mac OS X

Changes to these modules will still need to undergo normal backwards
compatibility assessments, but otherwise they will be kept entirely aligned
with the latest feature release of their Python 3 counterparts. This is
designed to make it easier to implement secure networked software in
Python, even for software that currently needs to remain compatible with
the Python 2 series (which includes a lot of network infrastructure
software).

The development branches will be arranged in such a way that even though
any further Python 2.7 releases published by the core development team
provide updated network security infrastructure, it will remain possible
for downstream redistributors to publish "Python 2.7 with legacy SSL
infrastructure" if they choose to do so.

While this PEP does not make any changes to the core development team's
handling of security-fix-only branches that are no longer in active
maintenance, it *does* recommend that commercial redistributors providing
extended support periods for the Python standard library either adopt a
similar approach to ensuring that the secure networking infrastructure
keeps pace with the evolution of the internet, or else disclaim support
for the use of older versions in roles that involving connecting
directly to the public internet.


Exemption Policy
================

Under this policy, the following network security related modules are
granted a blanket exemption to the restriction against adding new features
in maintenance releases, for the purpose of keeping their APIs aligned with
their counterparts in the latest feature release of Python 3:

* the ``ssl`` module
* the ``hashlib`` module
* the ``hmac`` module

This exemption applies to *all* proposals to backport backwards compatible
changes in these modules to Python 2.7 maintenance releases. This choice is
made deliberately to ensure that the "feature or fix?" argument isn't simply
replaced by a "security related or not?" argument. These particular modules
are inherently security related, and all enhancements to them improve
Python's capabilities as a platform for development of secure networked
software.

As part of this policy, permission is also granted to upgrade to newer
feature releases of OpenSSL when preparing the binary installers
for new maintenance releases of CPython.

Note that the ``sha`` and ``md5`` modules are not covered by this policy,
but have been deprecated in favour of ``hashlib`` since Python 2.5 and have
been removed entirely in the Python 3 series.

In addition to the above blanket exemption, a conditional exemption is
granted for these modules that may include some network security related
features:

* the ``os`` module (specifically ``os.urandom``)
* the ``random`` module

This more limited exemption for these modules requires that the *specific*
enhancement being proposed for backporting needs to be justified as being
network security related. This is generally restricted to cases where the
feature in question is needed by an update to one of the modules covered
by the blanket exemption.


Backwards Compatibility Considerations
======================================

This PEP does *not* grant any general exemptions to the usual backwards
compatibility policy for maintenance releases. Instead, by explicitly
encouraging the use of feature based checks and explicitly opting in to
less secure configurations, it is designed to make it easier to provide
more "secure by default" behaviour in future feature releases, while still
limiting the risk of breaking currently working software when upgrading to
a new maintenance release.

In *all* cases where this policy is applied to backport enhancements to
maintenance releases, it MUST be possible to write cross-version compatible
code that operates by "feature detection" (for example, checking for
particular attributes in the module), without needing to explicitly check
the Python version.

It is then up to library and framework code to provide an appropriate warning
and fallback behaviour if a desired feature is found to be missing. While
some especially security sensitive software MAY fail outright if a desired
security feature is unavailable, most software SHOULD instead emit a warning
and continue operating using a slightly degraded security configuration.

Affected APIs SHOULD be designed to allow library and application code to
perform the following actions after detecting the presence of a relevant
network security related feature:

* explicitly opt in to more secure settings (to allow the use of enhanced
  security features in older maintenance releases of Python)
* explicitly opt in to less secure settings (to allow the use of newer Python
  feature releases in lower security environments)
* determine the default setting for the feature (this MAY require explicit
  Python version checks to determine the Python feature release, but MUST
  NOT require checking for a specific maintenance release)

Security related changes to other modules (such as higher level networking
libraries and data format processing libraries) will continue to be made
available as backports and new modules on the Python Package Index, as
independent distribution remains the preferred approach to handling
software that must continue to evolve to handle changing development
requirements independently of the Python 2 standard library. Refer to
the `Motivation and Rationale`_ section for a review of the characteristics
that make the secure networking infrastructure worthy of special
consideration.


OpenSSL compatibility
---------------------

Under this policy, OpenSSL may be upgraded to more recent feature releases
in Python 2.7 maintenance releases. On Linux and most other POSIX systems,
the specific version of OpenSSL used already varies, as Python dynamically
links to the system provided OpenSSL library by default.

For the Windows binary installers, the ``_ssl`` and ``_hashlib`` modules are
statically linked with OpenSSL and the associated symbols are not exported.
Marc-Andre Lemburg indicates that updating to newer OpenSSL releases in the
``egenix-pyopenssl`` binaries has not resulted in any reported compatibility
issues [3]_

The Mac OS X binary installers historically followed the same policy as
other POSIX installations and dynamically linked to the Apple provided
OpenSSL libraries. However, Apple has now ceased updating these
cross-platform libraries, instead requiring that even cross-platform
developers adopt Mac OS X specific interfaces to access up to date security
infrastructure on their platform. Accordingly, and independently of this
PEP, the Mac OS X binary installers were already going to be switched to
statically linker newer versions of OpenSSL [4]_


Other Considerations
====================

Maintainability
---------------

This policy does NOT represent a commitment by volunteer contributors to
actually backport network security related changes from the Python 3 series
to the Python 2 series. Rather, it is intended to send a clear signal to
potential corporate contributors that the core development team are willing
to accept offers of corporate assistance in putting this policy into
effect and handling the resulting increase in the Python 2 maintenance
load.

Backporting security related fixes and enhancements to earlier versions is
a common service for commercial redistributors to offer to their customers.
This policy represents an explicit invitation to contribute some of those
changes back to the upstream community in cases where they are likely to
have a broad impact that helps improve the security of the internet as a
whole, rather than sitting back and waiting for unpaid volunteers to do it
for them.


Documentation
-------------

All modules covered by this policy MUST include a "Security Considerations"
section in their documentation.

In addition to any other module specific contents, this section MUST
enumerate key security enhancements and fixes (with CVE identifiers where
applicable), along with the feature and maintenance releases that first
included them.


Security releases
-----------------

This PEP does not propose any changes to the handling of security
releases - those will continue to be source only releases that
include only critical security fixes.

However, the recommendations for library and application developers are
deliberately designed to accommodate commercial redistributors that choose
to apply this policy to additional Python release series that are either in
security fix only mode, or have been declared "end of life" by the core
development team.

Whether or not redistributors choose to exercise that option will be up
to the individual redistributor.


Integration testing
-------------------

Third party integration testing services should offer users the ability
to test against both the latest Python 2.7 maintenance release and the
latest "Python 2.7 with legacy SSL infrastructure" release, to ensure that
libraries, frameworks and applications handle the legacy security
infrastructure correctly (either failing or degrading gracefully, depending
on the security sensitivity of the software).


Handling lower security environments with low risk tolerance
------------------------------------------------------------

For better or for worse (mostly worse), there are some environments where
the risk of latent security defects is more tolerated than the risk of
regressions in maintenance releases. This policy largely excludes these
environments from consideration where the modules covered by the exemption
are concerned.

However, one concession is made for the benefit of such environments: while
the main ``2.7`` branch in Mercurial will include the updated network
security infrastructure, the development process will be updated to include
a ``2.7-legacy-ssl`` branch with the more limited feature set of the
original 2.7 network security infrastructure, allowing downstream
redistributors to continue to provide Python 2.7 with the legacy SSL
infrastructure if they choose to do so. This branch will be tested
on the CPython continuous integration infrastructure, but no releases will
be made from it by the core development team.

As noted above, corporate redistributors and users are expected to provide
the additional development effort needed to make this practical. It is not
acceptable to expect volunteer contributors to resolve a problem created
largely by conservative corporate development practices.


Evolution of this Policy
========================

The key requirement for a module to be considered for inclusion in this
policy (whether under a blanket or conditional exemption) is that it must
have security implications *beyond* the specific application that is written
in Python and the system that application is running on. Thus the focus on
network security protocols and related cryptographic infrastructure - Python
is a popular choice for the development of web services and clients, and
thus the capabilities of widely used Python versions have implications for
the security design of other services that may themselves be using newer
versions of Python or other development languages, but need to interoperate
with clients or servers written using older versions of Python.

The intent behind this requirement is to minimise any impact that the
introduction of this policy may have on the stability and compatibility of
maintenance releases. It would be thoroughly counterproductive if end
users became as cautious about updating to new Python maintenance releases
as they are about updating to new feature releases.


Motivation and Rationale
========================

This PEP can be seen as a more targeted version of the "faster standard
library release cycle" proposals discussed in PEP 407 and PEP 413,
focusing specifically on those areas which have implications beyond the
Python community.


Background
----------

The creation of this PEP was prompted primarily by the aging SSL support in
the Python 2 series. As of March 2014, the Python 2.7 SSL module is
approaching four years of age, and the SSL support in the still popular
Python 2.6 release had its feature set locked six years ago.

These are simply too old to provide a foundation that can be recommended
in good conscience for secure networking software that operates over the
public internet, especially in an era where it is becoming quite clearly
evident that advanced persistent security threats are even more widespread
and more indiscriminate in their targeting than had previously been
understood. While they represented reasonable security infrastructure in
their time, the state of the art has moved on, and we need to investigate
mechanisms for effectively providing more up to date network security
infrastructure for users that, for whatever reason, are not currently in
a position to migrate to Python 3.

While the use of the system OpenSSL installation addresses many of these
concerns on Linux platforms, it doesn't address all of them, and in the
case of the binary installers for Windows and Mac OS X that are published
on python.org, the version of OpenSSL used is entirely within the control
of the Python core development team, and currently limited to OpenSSL
maintenance releases for the version initially shipped with the corresponding
Python feature release.

With increased popularity comes increased responsibility, and this policy
aims to acknowledge the fact that Python's popularity and adoption has now
reached a level where some of our design and policy decisions have
significant implications beyond the Python development community.

As one example, the Python 2 ``ssl`` module does not support the Server
Name Identification standard. While it is possible to obtain SNI support
by using the third party ``requests`` client library, actually doing so
currently requires using not only ``requests`` and its embedded dependencies,
but also half a dozen or more additional libraries. The lack of support
in the Python 2 series thus serves as an impediment to making effective
use of SNI on servers, as Python 2 clients will frequently fail to handle
it correctly.

Another more critical example is the lack of SSL hostname matching in the
Python 2 standard library - it is currently necessary to rely on a third
party library, such as ``requests`` or ``backports.ssl_match_hostname`` to
obtain that functionality in Python 2.

The Python 2 series also remains more vulnerable to remote timing attacks
on security sensitive comparisons than the Python 3 series, as it lacks a
standard library equivalent to the timing attack resistant
``hmac.compare_digest()`` function. While appropriate secure comparison
functions can be implemented in third party extensions, may users don't
even consider the issue and use ordinary equality comparisons instead
- while a standard library solution doesn't automatically fix that problem,
it *does* make the barrier to resolution much lower once the problem is
pointed out.

My position on the ongoing transition from Python 2 to Python 3 has long
been that Python 2 remains a supported platform for the core development
team, and that commercial support will remain available well after upstream
maintenance ends. However, in the absence of this network security
enhancement policy, that position is difficult to justify when it comes to
software that operates over the public internet. Just as many developers
consider it too difficult to develop truly secure modern networked software
in C/C++ (largely due to the challenges associated with manual
memory management), I anticipate that in the not too distant future, it
will be considered too difficult to develop truly secure modern networked
software using the Python 2 series (some developers would argue that we
have already reached that point).


Alternative: advise developers of networked software to migrate to Python 3
---------------------------------------------------------------------------

This alternative represents the status quo. Unfortunately, it has proven
to be unworkable in practice, as the backwards compatibility implications
mean that this is a non-trivial migration process for large applications
and integration projects. While the tools for migration have evolved to
a point where it is possible to migrate even large applications
opportunistically and incrementally (rather than all at once) by updating
code to run in the large common subset of Python 2 and Python 3, using the
most recent technology often isn't a priority in commercial environments.

Previously, this was considered an acceptable harm, as while it was an
unfortunate problem for the affected developers to have to face, it was
seen as an issue between them and their management chain to make the case
for infrastructure modernisation, and this case would become naturally
more compelling as the Python 3 series evolved.

However, now that we're fully aware of the impact the limitations of the
Python 2 standard library may be having on the evolution of internet
security standards, I no longer believe that it is reasonable to expect
platform and application developers to resolve all of the latent defects
in an application's Unicode correctness solely in order to gain access to
the network security enhancements already available in Python 3.

While Ubuntu (and to some extent Debian as well) are committed to porting all
default system services and scripts to Python 3, and to removing Python 2
from its default distribution images (but not from its archives), this is
a mammoth task and won't be completed for the Ubuntu 14.04 LTS release
(at least for the desktop image - it may be achieved for the mobile and
server images).

Fedora has even more work to do to migrate, and it will take a non-trivial
amount of time to migrate the relevant infrastructure components. While
Red Hat are also actively working to make it easier for users to use more
recent versions of Python on our stable platforms, it's going to take time
for those efforts to start having an impact on end users' choice of version,
and any such changes also don't benefit the core platform infrastructure
that runs in the integrated system Python by necessity.

The OpenStack migration to Python 3 is also still in its infancy, and even
though that's a project with an extensive and relatively robust automated
test suite, it's still large enough that it is going to take quite some time
to migrate.

And that's just three of the highest profile open source projects that
make heavy use of Python. Given the likely existence of large amounts of
legacy code that lacks the kind of automated regression test suite needed
to help support a migration from Python 2 to Python 3, there are likely to
be many cases where reimplementation (perhaps even in Python 3) proves
easier than migration. The key point of this PEP is that those situations
affect more people than just the developers and users of the affected
application: the existence of clients and servers with outdated network
security infrastructure becomes something that developers of secure
networked services need to take into account as part of their security
design, and that's a problem that inhibits the adoption of better security
standards.

As Terry Reedy noted, if we try to persist with the status quo, the likely
outcome is that commercial redistributors will attempt to do something
like this on behalf of their customers *anyway*, but in a potentially
inconsistent and ad hoc manner. By drawing the scope definition process
into the upstream project we are in a better position to influence the
approach taken to address the situation and to help ensure some consistency
across redistributors.

The problem is real, so *something* needs to change, and this PEP describes
my currently preferred approach to addressing the situation.


Alternative: create and release Python 2.8
------------------------------------------

With sufficient corporate support, it likely *would* be possible to create
and release Python 2.8 (it's highly unlikely such a project would garner
enough interest to be achievable with only volunteers). However, this
wouldn't actually solve the problem, as the aim is to provide a *relatively
low impact* way to incorporate enhanced security features into integrated
products and deployments that make use of Python 2. Upgrading to a new
Python feature release would mean both more work for the core development
team, as well as a more disruptive update that most potential end users
would likely just skip entirely.

Attempting to create a Python 2.8 release would also bring in suggestions
to backport many additional features from Python 3 (such as ``tracemalloc``
and the improved coroutine support), making the migration from Python 2.7
to this hypothetical 2.8 release even riskier and more disruptive.

This is not a recommended approach, as it would involve substantial
additional work for a result that is actually less effective in achieving
the original aim (which is to eliminate the current widespread use of the
aging network security infrastructure in the Python 2 series).


Alternative: distribute the security enhancements via PyPI
----------------------------------------------------------

While this initially appears to be an attractive and easier to manage
approach, it actually suffers from several significant problems.

Firstly, this is complex, low level, cross-platform code that integrates
with the underlying operating system across a variety of POSIX platforms
(including Mac OS X) and Windows. The CPython BuildBot fleet is already set
up to handle continuous integration in that context, but most of the
freely available continuous integration services just offer Linux, and
perhaps paid access to Windows. Those services work reasonably well for
software that largely runs on the abstraction layers offered by Python and
other dynamic languages, as well as the more comprehensive abstraction
offered by the JVM, but won't suffice for the kind of code involved here.

The OpenSSL dependency for the network security support also qualifies as
the kind of "complex binary dependency" that isn't yet handled well by the
``pip`` based software distribution ecosystem. Relying on a third party
binary dependency also creates potential compatibility problems for ``pip``
when running on other interpreters like ``PyPy``.

Another practical problem with the idea is the fact that ``pip`` itself
relies on the ``ssl`` support in the standard library (with some additional
support from a bundled copy of ``requests``, which in turn bundles
``backport.ssl_match_hostname``), and hence would require any replacement
module to also be bundled within ``pip``. This wouldn't pose any
insurmountable difficulties (it's just another dependency to vendor), but
it *would* mean yet another copy of OpenSSL to keep up to date.

This approach also has the same flaw as all other "improve security by
renaming things" approaches: they completely miss the users who most need
help, and raise significant barriers against being able to encourage users
to do the right thing when their infrastructure supports it (since
"use this other module" is a much higher impact change than "turn on this
higher security setting"). Deprecating the aging SSL infrastructure in the
standard library in favour of an external module would be even more user
hostile than accepting the slightly increased risk of regressions associated
with upgrading it in place.

Last, but certainly not least, this approach suffers from the same problem
as the idea of doing a Python 2.8 release: likely not solving the actual
problem. Commercial redistributors of Python are set up to redistribute
*Python*, and a pre-existing set of additional packages. Getting new
packages added to the pre-existing set *can* be done, but means approaching
each and every redistributor and asking them to update their
repackaging process accordingly. By contrast, the approach described in
this PEP would require redistributors to deliberately *opt out* of the
security enhancements (by switching to redistributing directly from the
``2.7-legacy-ssl`` branch rather than the main ``2.7`` branch), which most
of them are unlikely to do.


Open Questions
==============

* I believe I've addressed all the technical and scope questions I had, or
  others raised. That just leaves the question of "If we agree to this plan,
  who is actually going to handle all the extra work involved?" :)


Disclosure of Interest
======================

The author of this PEP currently works for Red Hat on test automation tools.
If this proposal is accepted, I will be strongly encouraging Red Hat to take
advantage of the resulting opportunity to help improve the overall security
of the Python ecosystem. However, I do not speak for Red Hat in this matter,
and cannot make any commitments on Red Hat's behalf.


Acknowledgements
================

Thanks to Christian Heimes for his recent efforts on greatly improving
Python's SSL support in the Python 3 series, and a variety of members of
the Python community for helping me to better understand the implications
of the default settings we provide in our SSL modules, and the impact that
tolerating the use of SSL infrastructure that was defined in 2010
(Python 2.7) or even 2008 (Python 2.6) potentially has for the security
of the web as a whole.

Christian and Donald Stufft also provided valuable feedback on a preliminary
draft of this proposal.

Thanks also to participants in the python-dev mailing list threads [1,2]_


References
==========

.. [1] https://mail.python.org/pipermail/python-dev/2014-March/133334.html
.. [2] https://mail.python.org/pipermail/python-dev/2014-March/133389.html
.. [3] https://mail.python.org/pipermail/python-dev/2014-March/133438.html
.. [4] https://mail.python.org/pipermail/python-dev/2014-March/133347.html


Copyright
=========

This document has been placed in the public domain.

-- 
Nick Coghlan   |   ncoghlan at gmail.com   |   Brisbane, Australia