python-peps/pep-0506.txt

PEP: 506
Title: Adding A Secrets Module To The Standard Library
Version: $Revision$
Last-Modified: $Date$
Author: Steven D'Aprano <steve@pearwood.info>
Status: Final
Type: Standards Track
Content-Type: text/x-rst
Created: 19-Sep-2015
Python-Version: 3.6
Post-History:


Abstract
========

This PEP proposes the addition of a module for common security-related
functions such as generating tokens to the Python standard library.


Definitions
===========

Some common abbreviations used in this proposal:

* PRNG:

  Pseudo Random Number Generator.  A deterministic algorithm used
  to produce random-looking numbers with certain desirable
  statistical properties.

* CSPRNG:

  Cryptographically Strong Pseudo Random Number Generator.  An
  algorithm used to produce random-looking numbers which are
  resistant to prediction.

* MT:

  Mersenne Twister.  An extensively studied PRNG which is currently
  used by the ``random`` module as the default.


Rationale
=========

This proposal is motivated by concerns that Python's standard library
makes it too easy for developers to inadvertently make serious security
errors.  Theo de Raadt, the founder of OpenBSD, contacted Guido van Rossum
and expressed some concern [#]_ about the use of MT for generating sensitive
information such as passwords, secure tokens, session keys and similar.

Although the documentation for the ``random`` module explicitly states that
the default is not suitable for security purposes [#]_, it is strongly
believed that this warning may be missed, ignored or misunderstood by
many Python developers.  In particular:

* developers may not have read the documentation and consequently
  not seen the warning;

* they may not realise that their specific use of the module has security
  implications; or

* not realising that there could be a problem, they have copied code
  (or learned techniques) from websites which don't offer best
  practises.

The first [#]_ hit when searching for "python how to generate passwords" on
Google is a tutorial that uses the default functions from the ``random``
module [#]_.  Although it is not intended for use in web applications, it is
likely that similar techniques find themselves used in that situation.
The second hit is to a StackOverflow question about generating
passwords [#]_.  Most of the answers given, including the accepted one, use
the default functions.  When one user warned that the default could be
easily compromised, they were told "I think you worry too much." [#]_

This strongly suggests that the existing ``random`` module is an attractive
nuisance when it comes to generating (for example) passwords or secure
tokens.

Additional motivation (of a more philosophical bent) can be found in the
post which first proposed this idea [#]_.


Proposal
========

Alternative proposals have focused on the default PRNG in the ``random``
module, with the aim of providing "secure by default" cryptographically
strong primitives that developers can build upon without thinking about
security.  (See Alternatives below.)  This proposes a different approach:

* The standard library already provides cryptographically strong
  primitives, but many users don't know they exist or when to use them.

* Instead of requiring crypto-naive users to write secure code, the
  standard library should include a set of ready-to-use "batteries" for
  the most common needs, such as generating secure tokens.  This code
  will both directly satisfy a need ("How do I generate a password reset
  token?"), and act as an example of acceptable practises which
  developers can learn from [#]_.

To do this, this PEP proposes that we add a new module to the standard
library, with the suggested name ``secrets``.  This module will contain a
set of ready-to-use functions for common activities with security
implications, together with some lower-level primitives.

The suggestion is that ``secrets`` becomes the go-to module for dealing
with anything which should remain secret (passwords, tokens, etc.)
while the ``random`` module remains backward-compatible.


API and Implementation
======================

This PEP proposes the following functions for the ``secrets`` module:

* Functions for generating tokens suitable for use in (e.g.) password
  recovery, as session keys, etc., in the following formats:

  - as bytes, ``secrets.token_bytes``;
  - as text, using hexadecimal digits, ``secrets.token_hex``;
  - as text, using URL-safe base-64 encoding, ``secrets.token_urlsafe``.

* A limited interface to the system CSPRNG, using either ``os.urandom``
  directly or ``random.SystemRandom``.  Unlike the ``random`` module, this
  does not need to provide methods for seeding, getting or setting the
  state, or any non-uniform distributions.  It should provide the
  following:

  - A function for choosing items from a sequence, ``secrets.choice``.
  - A function for generating a given number of random bits and/or bytes
    as an integer, ``secrets.randbits``.
  - A function for returning a random integer in the half-open range
    0 to the given upper limit, ``secrets.randbelow`` [#]_.

* A function for comparing text or bytes digests for equality while being
  resistant to timing attacks, ``secrets.compare_digest``.

The consensus appears to be that there is no need to add a new CSPRNG to
the ``random`` module to support these uses, ``SystemRandom`` will be
sufficient.

Some illustrative implementations have been given by Nick Coghlan [#]_
and a minimalist API by Tim Peters [#]_. This idea has also been discussed
on the issue tracker for the "cryptography" module [#]_.  The following
pseudo-code should be taken as the starting point for the real
implementation::

    from random import SystemRandom
    from hmac import compare_digest

    _sysrand = SystemRandom()

    randbits = _sysrand.getrandbits
    choice = _sysrand.choice

    def randbelow(exclusive_upper_bound):
        return _sysrand._randbelow(exclusive_upper_bound)

    DEFAULT_ENTROPY = 32  # bytes

    def token_bytes(nbytes=None):
        if nbytes is None:
            nbytes = DEFAULT_ENTROPY
        return os.urandom(nbytes)

    def token_hex(nbytes=None):
        return binascii.hexlify(token_bytes(nbytes)).decode('ascii')

    def token_urlsafe(nbytes=None):
        tok = token_bytes(nbytes)
        return base64.urlsafe_b64encode(tok).rstrip(b'=').decode('ascii')


The ``secrets`` module itself will be pure Python, and other Python
implementations can easily make use of it unchanged, or adapt it as
necessary. An implementation can be found on BitBucket [#]_.

Default arguments
~~~~~~~~~~~~~~~~~

One difficult question is "How many bytes should my token be?".  We can
help with this question by providing a default amount of entropy for the
"token_*" functions. If the ``nbytes`` argument is None or not given, the
default entropy will be used. This default value should be large enough
to be expected to be secure for medium-security uses, but is expected to
change in the future, possibly even in a maintenance release [#]_.

Naming conventions
~~~~~~~~~~~~~~~~~~

One question is the naming conventions used in the module [#]_, whether to
use C-like naming conventions such as "randrange" or more Pythonic names
such as "random_range".

Functions which are simply bound methods of the private ``SystemRandom``
instance (e.g. ``randrange``), or a thin wrapper around such, should keep
the familiar names. Those which are something new (such as the various
``token_*`` functions) will use more Pythonic names.

Alternatives
============

One alternative is to change the default PRNG provided by the ``random``
module [#]_.  This received considerable scepticism and outright opposition:

* There is fear that a CSPRNG may be slower than the current PRNG (which
  in the case of MT is already quite slow).

* Some applications (such as scientific simulations, and replaying
  gameplay) require the ability to seed the PRNG into a known state,
  which a CSPRNG lacks by design.

* Another major use of the ``random`` module is for simple "guess a number"
  games written by beginners, and many people are loath to make any
  change to the ``random`` module which may make that harder.

* Although there is no proposal to remove MT from the ``random`` module,
  there was considerable hostility to the idea of having to opt-in to
  a non-CSPRNG or any backwards-incompatible changes.

* Demonstrated attacks against MT are typically against PHP applications.
  It is believed that PHP's version of MT is a significantly softer target
  than Python's version, due to a poor seeding technique [#]_.  Consequently,
  without a proven attack against Python applications, many people object
  to a backwards-incompatible change.

Nick Coghlan made an :pep:`earlier suggestion <504>`
for a globally configurable PRNG
which uses the system CSPRNG by default, but has since withdrawn it
in favour of this proposal.


Comparison To Other Languages
=============================

* PHP

  PHP includes a function ``uniqid`` [#]_ which by default returns a
  thirteen character string based on the current time in microseconds.
  Translated into Python syntax, it has the following signature::

    def uniqid(prefix='', more_entropy=False)->str

  The PHP documentation warns that this function is not suitable for
  security purposes.  Nevertheless, various mature, well-known PHP
  applications use it for that purpose (citation needed).

  PHP 5.3 and better also includes a function ``openssl_random_pseudo_bytes``
  [#]_.  Translated into Python syntax, it has roughly the following
  signature::

    def openssl_random_pseudo_bytes(length:int)->Tuple[str, bool]

  This function returns a pseudo-random string of bytes of the given
  length, and a boolean flag giving whether the string is considered
  cryptographically strong.  The PHP manual suggests that returning
  anything but True should be rare except for old or broken platforms.

* JavaScript

  Based on a rather cursory search [#]_, there do not appear to be any
  well-known standard functions for producing strong random values in
  JavaScript. ``Math.random`` is often used, despite serious weaknesses
  making it unsuitable for cryptographic purposes [#]_. In recent years
  the majority of browsers have gained support for ``window.crypto.getRandomValues`` [#]_.

  Node.js offers a rich cryptographic module, ``crypto`` [#]_, most of
  which is beyond the scope of this PEP. It does include a single function
  for generating random bytes, ``crypto.randomBytes``.

* Ruby

  The Ruby standard library includes a module ``SecureRandom`` [#]_
  which includes the following methods:

  * base64 - returns a Base64 encoded random string.

  * hex - returns a random hexadecimal string.

  * random_bytes - returns a random byte string.

  * random_number - depending on the argument, returns either a random
    integer in the range(0, n), or a random float between 0.0 and 1.0.

  * urlsafe_base64 - returns a random URL-safe Base64 encoded string.

  * uuid - return a version 4 random Universally Unique IDentifier.


What Should Be The Name Of The Module?
======================================

There was a proposal to add a "random.safe" submodule, quoting the Zen
of Python "Namespaces are one honking great idea" koan.  However, the
author of the Zen, Tim Peters, has come out against this idea [#]_, and
recommends a top-level module.

In discussion on the python-ideas mailing list so far, the name "secrets"
has received some approval, and no strong opposition.

There is already an existing third-party module with the same name [#]_,
but it appears to be unused and abandoned.


Frequently Asked Questions
==========================

* Q: Is this a real problem? Surely MT is random enough that nobody can
  predict its output.

  A: The consensus among security professionals is that MT is not safe
  in security contexts.  It is not difficult to reconstruct the internal
  state of MT [#]_ [#]_ and so predict all past and future values.  There
  are a number of known, practical attacks on systems using MT for
  randomness [#]_.

* Q: Attacks on PHP are one thing, but are there any known attacks on
  Python software?

  A: Yes.  There have been vulnerabilities in Zope and Plone at the very
  least.  Hanno Schlichting commented [#]_::

      "In the context of Plone and Zope a practical attack was
      demonstrated, but I can't find any good non-broken links about
      this anymore.  IIRC Plone generated a random number and exposed
      this on each error page along the lines of 'Sorry, you encountered
      an error, your problem has been filed as <random number>, please
      include this when you contact us'.  This allowed anyone to do large
      numbers of requests to this page and get enough random values to
      reconstruct the MT state.  A couple of security related modules used
      random instead of system random (cookie session ids, password reset
      links, auth token), so the attacker could break all of those."

  Christian Heimes reported this issue to the Zope security team in 2012 [#]_,
  there are at least two related CVE vulnerabilities [#]_, and at least one
  work-around for this issue in Django [#]_.

* Q: Is this an alternative to specialist cryptographic software such as SSL?

  A: No. This is a "batteries included" solution, not a full-featured
  "nuclear reactor".  It is intended to mitigate against some basic
  security errors, not be a solution to all security-related issues. To
  quote Nick Coghlan referring to his earlier proposal [#]_::

      "...folks really are better off learning to use things like
      cryptography.io for security sensitive software, so this change
      is just about harm mitigation given that it's inevitable that a
      non-trivial proportion of the millions of current and future
      Python developers won't do that."

* Q: What about a password generator?

  A: The consensus is that the requirements for password generators are too
  variable for it to be a good match for the standard library [#]_. No password
  generator will be included in the initial release of the module, instead it
  will be given in the documentation as a recipe (à la the recipes in the
  ``itertools`` module) [#]_.

* Q: Will ``secrets`` use /dev/random (which blocks) or /dev/urandom (which
  doesn't block) on Linux? What about other platforms?

  A: ``secrets`` will be based on ``os.urandom`` and ``random.SystemRandom``,
  which are interfaces to your operating system's best source of cryptographic
  randomness. On Linux, that may be ``/dev/urandom`` [#]_, on Windows it may be
  ``CryptGenRandom()``, but see the documentation and/or source code for the
  detailed implementation details.


References
==========

.. [#] https://mail.python.org/pipermail/python-ideas/2015-September/035820.html

.. [#] https://docs.python.org/3/library/random.html

.. [#] As of the date of writing. Also, as Google search terms may be
       automatically customised for the user without their knowledge, some
       readers may see different results.

.. [#] http://interactivepython.org/runestone/static/everyday/2013/01/3_password.html

.. [#] http://stackoverflow.com/questions/3854692/generate-password-in-python

.. [#] http://stackoverflow.com/questions/3854692/generate-password-in-python/3854766#3854766

.. [#] https://mail.python.org/pipermail/python-ideas/2015-September/036238.html

.. [#] At least those who are motivated to read the source code and documentation.

.. [#] After considerable discussion, Guido ruled that the module need only
       provide ``randbelow``, and not similar functions ``randrange`` or
       ``randint``.  http://code.activestate.com/lists/python-dev/138375/

.. [#] https://mail.python.org/pipermail/python-ideas/2015-September/036271.html

.. [#] https://mail.python.org/pipermail/python-ideas/2015-September/036350.html

.. [#] https://github.com/pyca/cryptography/issues/2347

.. [#] https://bitbucket.org/sdaprano/secrets

.. [#] https://mail.python.org/pipermail/python-ideas/2015-September/036517.html
       https://mail.python.org/pipermail/python-ideas/2015-September/036515.html

.. [#] https://mail.python.org/pipermail/python-ideas/2015-September/036474.html

.. [#] Link needed.

.. [#] By default PHP seeds the MT PRNG with the time (citation needed),
       which is exploitable by attackers, while Python seeds the PRNG with
       output from the system CSPRNG, which is believed to be much harder to
       exploit.

.. [#] http://php.net/manual/en/function.uniqid.php

.. [#] http://php.net/manual/en/function.openssl-random-pseudo-bytes.php

.. [#] Volunteers and patches are welcome.

.. [#] http://ifsec.blogspot.fr/2012/05/cross-domain-mathrandom-prediction.html

.. [#] https://developer.mozilla.org/en-US/docs/Web/API/RandomSource/getRandomValues

.. [#] https://nodejs.org/api/crypto.html

.. [#] http://ruby-doc.org/stdlib-2.1.2/libdoc/securerandom/rdoc/SecureRandom.html

.. [#] https://mail.python.org/pipermail/python-ideas/2015-September/036254.html

.. [#] https://pypi.python.org/pypi/secrets

.. [#] https://jazzy.id.au/2010/09/22/cracking_random_number_generators_part_3.html

.. [#] https://mail.python.org/pipermail/python-ideas/2015-September/036077.html

.. [#] https://media.blackhat.com/bh-us-12/Briefings/Argyros/BH_US_12_Argyros_PRNG_WP.pdf

.. [#] Personal communication, 2016-08-24.

.. [#] https://bugs.launchpad.net/zope2/+bug/1071067

.. [#] http://www.cvedetails.com/cve/CVE-2012-5508/
       http://www.cvedetails.com/cve/CVE-2012-6661/

.. [#] https://github.com/django/django/commit/1525874238fd705ec17a066291935a9316bd3044

.. [#] https://mail.python.org/pipermail/python-ideas/2015-September/036157.html

.. [#] https://mail.python.org/pipermail/python-ideas/2015-September/036476.html
       https://mail.python.org/pipermail/python-ideas/2015-September/036478.html

.. [#] https://mail.python.org/pipermail/python-ideas/2015-September/036488.html

.. [#] http://sockpuppet.org/blog/2014/02/25/safely-generate-random-numbers/
       http://www.2uo.de/myths-about-urandom/


Copyright
=========

This document has been placed in the public domain.



..
   Local Variables:
   mode: indented-text
   indent-tabs-mode: nil
   sentence-end-double-space: t
   fill-column: 70
   coding: utf-8
   End: