python-peps/peps/pep-0740.rst

PEP: 740
Title: Index support for digital attestations
Author: William Woodruff <william@yossarian.net>,
        Facundo Tuesca <facundo.tuesca@trailofbits.com>,
        Dustin Ingram <di@python.org>
Sponsor: Donald Stufft <donald@stufft.io>
PEP-Delegate: Donald Stufft <donald@stufft.io>
Discussions-To: https://discuss.python.org/t/pep-740-index-support-for-digital-attestations/44498
Status: Draft
Type: Informational
Topic: Packaging
Created: 08-Jan-2024
Post-History: `02-Jan-2024 <https://discuss.python.org/t/pre-pep-exposing-trusted-publisher-provenance-on-pypi/42337>`__,
              `29-Jan-2024 <https://discuss.python.org/t/pep-740-index-support-for-digital-attestations/44498>`__

Abstract
========

This PEP proposes a collection of changes related to the upload and distribution
of digitally signed attestations and metadata used to verify them on a Python
package repository, such as PyPI.

These changes have two subcomponents:

* Changes to the currently unstandardized PyPI upload API, allowing clients
  to upload digital attestations as :ref:`attestation objects <attestation-object>`;
* Changes to the :pep:`503` and :pep:`691` "simple" APIs, allowing clients
  to retrieve both digital attestations and
  `Trusted Publishing <https://docs.pypi.org/trusted-publishers/>`_ metadata
  for individual release files as :ref:`provenance objects <provenance-object>`.

This PEP does not make a policy recommendation around mandatory digital
attestations on release uploads or their subsequent verification by installing
clients like ``pip``.

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

Desire for digital signatures on Python packages has been repeatedly
expressed by both package maintainers and downstream users:

* Maintainers wish to demonstrate the integrity and authenticity of their
  package uploads;
* Individual downstream users wish to verify package integrity and authenticity
  without placing additional trust in their index's honesty;
* "Bulk" downstream users (such as Operating System distributions) wish to
  perform similar verifications and potentially re-expose or countersign
  for their own downstream packaging ecosystems.

This proposal seeks to accommodate each of the above use cases.

Additionally, this proposal identifies the following motivations:

* Verifiable provenance for Python package distributions: many Python
  packages currently contain *unauthenticated* provenance metadata, such
  as URLs for source hosts. A cryptographic attestation format could enable
  strong *authenticated* links between these packages and their source hosts,
  allowing both the index and downstream users to cryptographically verify that
  a package originates from its claimed source repository.
* Raising attacker requirements: an attacker who seeks to take
  over a Python package can be described along *sophistication*
  (unsophisticated to sophisticated) and *targeting* dimensions
  (opportunistic to targeted).

  Digital attestations impose additional sophistication requirements: the
  attacker must be sufficiently sophisticated to access private signing material
  (or signing identities).
* Index verifiability: in the status quo, the only attestation provided by the
  index is an optional PGP signature per release file
  (see :ref:`PGP signatures <pgp-signatures>`). These signatures are not
  (and cannot be) checked by the index either for well-formedness or for
  validity, since the index has no mechanism for identifying the right public
  key for the signature. This PEP overcomes this limitation
  by ensuring that :ref:`provenance objects <provenance-object>` contain all
  of the metadata needed by the index to verify an attestation's validity.

This PEP proposes a generic attestation format, containing an
:ref:`attestation payload for signature generation <payload-and-signature-generation>`,
with the expectation that index providers adopt the
format with a suitable source of identity for signature verification, such as
Trusted Publishing.

Design Considerations
---------------------

This PEP identifies the following design considerations when evaluating
both its own proposed changes and previous work in the same or adjacent
areas of Python packaging:

1. Index accessibility: digital attestations for Python packages
   are ideally retrievable directly from the index itself, as "detached"
   resources.

   This both simplifies some compatibility concerns (by avoiding
   the need to modify the distribution formats themselves) and also simplifies
   the behavior of potential installing clients (by allowing them to
   retrieve each attestation before its corresponding package without needing
   to do streaming decompression).

2. Verification by the index itself: in addition to enabling verification
   by installing clients, each digital attestation is *ideally* verifiable
   in some form by the index itself.

   This both increases the overall quality
   of attestations uploaded to the index (preventing, for example, users
   from accidentally uploading incorrect or invalid attestations) and also
   enables UI and UX refinements on the index itself (such as a "provenance"
   view for each uploaded package).

3. General applicability: digital attestations should be applicable to
   *any and every* package uploaded to the index, regardless of its format
   (sdist or wheel) or interior contents.

4. Metadata support: this PEP refers to "digital attestations" rather than
   just "digital signatures" to emphasize the ideal presence of additional
   metadata within the cryptographic envelope.

   For example, to prevent domain separation between a distribution's name and
   its contents, this PEP proposes that digital attestations be performed over
   ``HASH(name || HASH(contents))`` rather than just ``HASH(contents)``.


Previous Work
-------------

.. _pgp-signatures:

PGP signatures
^^^^^^^^^^^^^^

PyPI and other indices have historically supported PGP signatures on uploaded
distributions. These could be supplied during upload, and could be retrieved
by installing clients via the ``data-gpg-sig`` attribute in the :pep:`503`
API, the ``gpg-sig`` key on the :pep:`691` API, or via an adjacent
``.asc``-suffixed URL.

PGP signature uploads have been disabled on PyPI since
`May 2023 <https://blog.pypi.org/posts/2023-05-23-removing-pgp/>`_, after
`an investigation <https://blog.yossarian.net/2023/05/21/PGP-signatures-on-PyPI-worse-than-useless>`_
determined that the majority of signatures (which, themselves, constituted a
tiny percentage of overall uploads) could not be associated with a public key or
otherwise meaningfully verified.

In their previously supported form on PyPI, PGP signatures satisfied
considerations (1) and (3) above but not (2) (owing to the need for external
keyservers and key distribution) or (4) (due to PGP signatures typically being
constructed over just an input file, without any associated signed metadata).

Wheel signatures
^^^^^^^^^^^^^^^^

:pep:`427` (and its :ref:`living PyPA counterpart <packaging:binary-distribution-format>`)
specify the :term:`wheel format <packaging:Wheel>`.

This format includes accommodations for digital signatures embedded directly
into the wheel, in either JWS or S/MIME format. These signatures are specified
over a :pep:`376` RECORD, which is modified to include a cryptographic digest
for each recorded file in the wheel.

While wheel signatures are fully specified, they do not appear to be broadly
used; the official `wheel tooling <https://github.com/pypa/wheel>`_ deprecated
signature generation and verification support
`in 0.32.0 <https://wheel.readthedocs.io/en/stable/news.html>`_, which was
released in 2018.

Additionally, wheel signatures do not satisfy any of
the above considerations (due to the "attached" nature of the signatures,
non-verifiability on the index itself, and support for wheels only).

Specification
=============

.. _upload-endpoint:

Upload endpoint changes
-----------------------

The current upload API is not standardized. However, we propose the following
changes to it:

* In addition to the current top-level ``content`` and ``gpg_signature`` fields,
  the index **SHALL** accept ``attestations`` as an additional multipart form
  field.
* The new ``attestations`` field **SHALL** be a JSON array.
* The ``attestations`` array **SHALL** have one or more items, each a JSON object
  representing an individual attestation.
* Each attestation object **MUST** be verifiable by the index. If the index fails
  to verify any attestation in ``attestations``, it **MUST** reject the upload.
  The format of attestation objects is defined under :ref:`attestation-object`
  and the process for verifying attestations is defined under
  :ref:`attestation-verification`.

Index changes
-------------

Simple Index
^^^^^^^^^^^^

* When an uploaded file has one or more attestations, the index **MAY**
  provide a ``.provenance`` file adjacent to the hosted distribution.
  The format of the ``.provenance`` file **SHALL** be a JSON-encoded
  :ref:`provenance object <provenance-object>`, which **SHALL** contain
  the file's attestations.

  For example, if an uploaded file is hosted at
  the URL ``https://example.com/sampleproject-1.2.3.tar.gz``, the provenance
  URL would be ``https://example.com/sampleproject-1.2.3.tar.gz.provenance``.

* When a ``.provenance`` file is present, the index **MAY** include a
  ``data-provenance`` attribute on its file link. The value of the
  ``data-provenance`` attribute **SHALL** be the SHA256 digest of the
  associated ``.provenance`` file.

* The index **MAY** choose to modify the ``.provenance`` file. For example,
  the index **MAY** permit adding additional attestations and verification
  materials, such as attestations from third-party auditors or other services.
  When the index modifies the ``.provenance`` file, it **MUST** also update the
  ``data-provenance`` attribute's value to the new SHA256 digest.

  See :ref:`changes-to-provenance-objects` for an additional discussion of
  reasons why a file's provenance may change.

JSON-based Simple API
^^^^^^^^^^^^^^^^^^^^^

* When an uploaded file has one or more attestations, the index **MAY**
  include a ``provenance`` object in the ``file`` dictionary for that file.
  The format of the ``provenance`` object **SHALL** be a JSON-encoded
  :ref:`provenance object <provenance-object>`, which **SHALL** contain
  the file's attestations.

* The index **MAY** choose to modify the ``provenance`` object, under the same
  conditions as the ``.provenance`` file specified above.

  See :ref:`changes-to-provenance-objects` for an additional discussion of
  reasons why a file's provenance may change.

These changes require a version change to the JSON API:

* The ``api-version`` **SHALL** specify version 1.2 or later.

.. _attestation-object:

Attestation objects
-------------------

An attestation object is a JSON object with several required keys; applications
or signers may include additional keys so long as all explicitly
listed keys are provided. The required layout of an attestation
object is provided as pseudocode below.

.. code-block:: python

  @dataclass
  class Attestation:
      version: Literal[1]
      """
      The attestation object's version, which is always 1.
      """

      verification_material: VerificationMaterial
      """
      Cryptographic materials used to verify `message_signature`.
      """

      message_signature: str
      """
      The attestation's signature, as `base64(raw-sig)`, where `raw-sig`
      is the raw bytes of the signing operation over the attestation payload.
      """

  @dataclass
  class VerificationMaterial:
      certificate: str
      """
      The signing certificate, as `base64(DER(cert))`.
      """

      transparency_entries: list[object]
      """
      One or more transparency log entries for this attestation's signature
      and certificate.
      """

A full data model for each object in ``transparency_entries`` is provided in
:ref:`appendix-2`. Attestation objects **SHOULD** include one or more
transparency log entries, and **MAY** include additional keys for other
sources of signed time (such as an :rfc:`3161` Time Stamping Authority or a
`Roughtime <https://blog.cloudflare.com/roughtime>`__ server).

Attestation objects are versioned; this PEP specifies version 1. Each version
is tied to a single cryptographic suite to minimize unnecessary cryptographic
agility. In version 1, the suite is as follows:

* Certificates are specified as X.509 certificates, and comply with the
  profile in :rfc:`5280`.
* The message signature algorithm is ECDSA, with the P-256 curve for public keys
  and SHA-256 as the cryptographic digest function.

Future PEPs may change this suite (and the overall shape of the attestation
object) by selecting a new version number.

.. _payload-and-signature-generation:

Attestation payload and signature generation
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

The *attestation payload* is the actual claim that is cryptographically signed
over within the attestation object (as the ``message_signature``).

The attestation payload is encoded as an :rfc:`8785` canonicalized JSON object,
with the following pseudocode layout:

.. code-block:: python

  @dataclass
  class AttestationPayload:
      distribution: str
      """
      The file name of the Python package distribution.
      """

      digest: str
      """
      The SHA-256 digest of the distribution's contents, as a hexadecimal string.
      """

The value of ``distribution`` is the same distribution filename that appears
in the :pep:`503` and :pep:`691` APIs. For example, ``distribution`` would be
``sampleproject-1.2.0-py2.py3-none-any.whl`` for the following simple index
entry:

.. code-block:: html

  <a href="https://example.com/...">sampleproject-1.2.0-py2.py3-none-any.whl</a><br/>

In practice, this means that ``distribution`` is defined by the PyPA's
living specifications for
:ref:`binary distributions <packaging:binary-distribution-format>` and
:ref:`source distributions <packaging:source-distribution-format>`, although
non-conforming distributions may be hosted by the index.

The following pseudocode demonstrates the construction of an attestation
payload and its signature:

.. code-block:: python

  def build_payload(dist: Path) -> AttestationPayload:
      return AttestationPayload(
        distribution=dist.name,
        digest=sha256(dist.read_bytes()).hexdigest,
      )

  attestation_payload = build_payload("sampleproject-1.2.0-py2.py3-none-any.whl")

  # canonical_json is a fictitious module that performs RFC 8785 canonical
  # JSON serialization.
  encoded_payload = canonical_json.dumps(asdict(attestation_payload))

  raw_signature = signing_key.sign(encoded_payload, ECDSA(SHA2_256()))
  message_signature = b64encode(raw_signature)

.. _provenance-object:

Provenance objects
------------------

The index will serve uploaded attestations along with metadata that can assist
in verifying them in the form of JSON serialized objects.

These *provenance objects* will be available via both the :pep:`503` Simple Index
and :pep:`691` JSON-based Simple API as described above, and will have the
following layout:

.. code-block:: json

    {
        "version": 1,
        "attestation_bundles": [
          {
            "publisher": {
              "kind": "important-ci-service",
              "claims": {},
              "vendor-property": "foo",
              "another-property": 123
            },
            "attestations": [
              { /* attestation 1 ... */ },
              { /* attestation 2 ... */ }
            ]
          }
        ]
    }

or, as pseudocode:

.. code-block:: python

  @dataclass
  class Publisher:
      kind: string
      """
      The kind of Trusted Publisher.
      """

      claims: object | None
      """
      Any context-specific claims retained by the index during Trusted Publisher
      authentication.
      """

      _rest: object
      """
      Each publisher object is open-ended, meaning that it MAY contain additional
      fields beyond the ones specified explicitly above. This field signals that,
      but is not itself present.
      """

  @dataclass
  class AttestationBundle:
      publisher: Publisher
      """
      The publisher associated with this set of attestations.
      """

      attestations: list[Attestation]
      """
      The set of attestations included in this bundle.
      """

  @dataclass
  class Provenance:
      version: Literal[1]
      """
      The provenance object's version, which is always 1.
      """

      attestation_bundles: list[AttestationBundle]
      """
      One or more attestation "bundles".
      """

* ``version`` is ``1``. Like attestation objects, provenance objects are
  versioned, and this PEP only defines version ``1``.
* ``attestation_bundles`` is a **required** JSON array, containing one
  or more "bundles" of attestations. Each bundle corresponds to a
  signing identity (such as a Trusted Publishing identity), and contains
  one or more attestation objects.

  As noted in the ``Publisher`` model,
  each ``AttestationBundle.publisher`` object is specific to its Trusted Publisher
  but must include at minimum:

  * A ``kind`` key, which **MUST** be a JSON string that uniquely identifies the
    kind of Trusted Publisher.
  * A ``claims`` key, which **MUST** be a JSON object containing any context-specific
    claims retained by the index during Trusted Publisher authentication.

  All other keys in the publisher object are publisher-specific. A full
  illustrative example of a publisher object is provided in :ref:`appendix-1`.

  Each array of attestation objects is a superset of the ``attestations``
  array supplied by the uploaded through the ``attestations`` field at upload
  time, as described in :ref:`upload-endpoint` and
  :ref:`changes-to-provenance-objects`.

.. _changes-to-provenance-objects:

Changes to provenance objects
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Provenance objects are *not* immutable, and may change over time. Reasons
for changes to the provenance object include but are not limited to:

* Addition of new attestations for a pre-existing signing identity: the index
  **MAY** choose to allow additional attestations by pre-existing signing
  identities, such as newer attestation versions for already uploaded
  files.

* Addition of new signing identities and associated attestations: the index
  **MAY** choose to support attestations from sources other than the file's
  uploader, such as third-party auditors or the index itself. These attestations
  may be performed asynchronously, requiring the index to insert them into
  the provenance object *post facto*.

.. _attestation-verification:

Attestation verification
------------------------

Verifying an attestation object requires verification of each of the following:

* ``version`` is ``1``. The verifier **MUST** reject any other version.
* ``verification_material.certificate`` is a valid signing certificate, as
  issued by an *a priori* trusted authority (such as a root of trust already
  present within the verifying client).
* ``verification_material.certificate`` identifies an appropriate signing
  subject, such as the machine identity of the Trusted Publisher that published
  the package.
* ``message_signature`` can be verified by ``verification_material.certificate``,
  using the reconstructed attestation payload as the cleartext input. The
  verifier **MUST** reconstruct the attestation payload itself.

In addition to the above required steps, a verifier **MAY** additionally verify
``verification_material.transparency_entries`` on a policy basis, e.g. requiring
at least one transparency log entry or a threshold of entries. When verifying
transparency entries, the verifier **MUST** confirm that the inclusion time for
each entry lies within the signing certificate's validity period.

Security Implications
=====================

This PEP is primarily "mechanical" in nature; it provides layouts for
structuring and serving verifiable digital attestations without specifying
higher level security "policies" around attestation validity, thresholds
between attestations, and so forth.

Cryptographic agility in attestations
-------------------------------------

Algorithmic agility is a common source of exploitable vulnerabilities
in cryptographic schemes. This PEP limits algorithmic agility in two ways:

* All algorithms are specified in a single suite, rather than a geometric
  collection of parameters. This makes it impossible (for example) for an
  attacker to select a strong signature algorithm with a weak hash function,
  compromising the scheme as a whole.
* Attestation objects are versioned, and may only contain the algorithmic
  suite specified for their version. If a specific suite
  is considered insecure in the future, clients may choose to blanket reject
  or qualify verifications of attestations that contain that suite.

Index trust
-----------

This PEP does **not** increase (or decrease) trust in the index itself:
the index is still effectively trusted to honestly deliver unmodified package
distributions, since a dishonest index capable of modifying package
contents could also dishonestly modify or omit package attestations.
As a result, this PEP's presumption of index trust is equivalent to the
unstated presumption with earlier mechanisms, like PGP and wheel signatures.

This PEP does not preclude or exclude future index trust mechanisms, such
as :pep:`458` and/or :pep:`480`.

Flexible attestations
---------------------

This PEP specifies a fixed attestation payload (defined in
:ref:`payload-and-signature-generation`), binding the contents of each uploaded
file to its public name on the index. This payload format is fixed and
inflexible to ease implementation, and to minimize additional mechanical
changes to the index itself (e.g., needing to store and service detached
attestation documents).

This PEP does not preclude or exclude future more flexible attestation payload
formats, such as ones built on `in-toto <https://in-toto.io/>`__.

Recommendations
===============

This PEP recommends, but does not mandate, that attestation objects
contain one or more verifiable sources of signed time that corroborate the
signing certificate's claimed validity period. Indices that implement this
PEP may choose to strictly enforce this requirement.

.. _appendix-1:

Appendix 1: Example Trusted Publisher Representation
====================================================

This appendix provides a fictional example of a ``publisher`` key within
a simple JSON API ``project.files[].provenance`` listing:

.. code-block:: json

    "publisher": {
        "kind": "GitHub",
        "claims": {
            "ref": "refs/tags/v1.0.0",
            "sha": "da39a3ee5e6b4b0d3255bfef95601890afd80709"
        },
        "repository_name": "HolyGrail",
        "repository_owner": "octocat",
        "repository_owner_id": "1",
        "workflow_filename": "publish.yml",
        "environment": null
    }


.. _appendix-2:

Appendix 2: Data models for Transparency Log Entries
====================================================

This appendix contains pseudocoded data models for transparency log entries
in attestation objects. Each transparency log entry serves as a source
of signed inclusion time, and can be verified either online or offline.

.. code-block:: python

  @dataclass
  class TransparencyLogEntry:
      log_index: int
      """
      The global index of the log entry, used when querying the log.
      """

      log_id: str
      """
      An opaque, unique identifier for the log.
      """

      entry_kind: str
      """
      The kind (type) of log entry.
      """

      entry_version: str
      """
      The version of the log entry's submitted format.
      """

      integrated_time: int
      """
      The UNIX timestamp from the log from when the entry was persisted.
      """

      inclusion_proof: InclusionProof
      """
      The actual inclusion proof the the log entry.
      """


  @dataclass
  class InclusionProof:
      log_index: int
      """
      The index of the entry in the tree it was written to.
      """

      root_hash: str
      """
      The digest stored at the root of the Merkle tree at the time of proof
      generation.
      """

      tree_size: int
      """
      The size of the Merkle tree at the time of proof generation.
      """

      hashes: list[str]
      """
      A list of hashes required to complete the inclusion proof, sorted
      in order from leaf to root. The leaf and root hashes are not themselves
      included in this list; the root is supplied via `root_hash` and the client
      must calculate the leaf hash.
      """

      checkpoint: str
      """
      The signed tree head's signature, at the time of proof generation.
      """

      cosigned_checkpoints: list[str]
      """
      Cosigned checkpoints from zero or more log witnesses.
      """

Copyright
=========

This document is placed in the public domain or under the
CC0-1.0-Universal license, whichever is more permissive.