PEP: 578
Title: Python Runtime Audit Hooks
Version: $Revision$
Last-Modified: $Date$
Author: Steve Dower <steve.dower@python.org>
BDFL-Delegate: Christian Heimes <christian@python.org>
Status: Accepted
Type: Standards Track
Content-Type: text/x-rst
Created: 16-Jun-2018
Python-Version: 3.8
Post-History: 28-Mar-2019, 07-May-2019
Abstract
========
This PEP describes additions to the Python API and specific behaviors
for the CPython implementation that make actions taken by the Python
runtime visible to auditing tools. Visibility into these actions
provides opportunities for test frameworks, logging frameworks, and
security tools to monitor and optionally limit actions taken by the
runtime.
This PEP proposes adding two APIs to provide insights into a running
Python application: one for arbitrary events, and another specific to
the module import system. The APIs are intended to be available in all
Python implementations, though the specific messages and values used
are unspecified here to allow implementations the freedom to determine
how best to provide information to their users. Some examples likely
to be used in CPython are provided for explanatory purposes.
See :pep:`551` for discussion and recommendations on enhancing the
security of a Python runtime making use of these auditing APIs.
Background
==========
Python provides access to a wide range of low-level functionality on
many common operating systems. While this is incredibly useful for
"write-once, run-anywhere" scripting, it also makes monitoring of
software written in Python difficult. Because Python uses native system
APIs directly, existing monitoring tools either suffer from limited
context or auditing bypass.
Limited context occurs when system monitoring can report that an
action occurred, but cannot explain the sequence of events leading to
it. For example, network monitoring at the OS level may be able to
report "listening started on port 5678", but may not be able to
provide the process ID, command line, parent process, or the local
state in the program at the point that triggered the action. Firewall
controls to prevent such an action are similarly limited, typically
to process names or some global state such as the current user, and
in any case rarely provide a useful log file correlated with other
application messages.
Auditing bypass can occur when the typical system tool used for an
action would ordinarily report its use, but accessing the APIs via
Python do not trigger this. For example, invoking "curl" to make HTTP
requests may be specifically monitored in an audited system, but
Python's "urlretrieve" function is not.
Within a long-running Python application, particularly one that
processes user-provided information such as a web app, there is a risk
of unexpected behavior. This may be due to bugs in the code, or
deliberately induced by a malicious user. In both cases, normal
application logging may be bypassed resulting in no indication that
anything out of the ordinary has occurred.
Additionally, and somewhat unique to Python, it is very easy to affect
the code that is run in an application by manipulating either the
import system's search path or placing files earlier on the path than
intended. This is often seen when developers create a script with the
same name as the module they intend to use - for example, a
``random.py`` file that attempts to import the standard library
``random`` module.
This is not sandboxing, as this proposal does not attempt to prevent
malicious behavior (though it enables some new options to do so).
See the `Why Not A Sandbox`_ section below for further discussion.
Overview of Changes
===================
The aim of these changes is to enable both application developers and
system administrators to integrate Python into their existing
monitoring systems without dictating how those systems look or behave.
We propose two API changes to enable this: an Audit Hook and Verified
Open Hook. Both are available from Python and native code, allowing
applications and frameworks written in pure Python code to take
advantage of the extra messages, while also allowing embedders or
system administrators to deploy builds of Python where auditing is
always enabled.
Only CPython is bound to provide the native APIs as described here.
Other implementations should provide the pure Python APIs, and
may provide native versions as appropriate for their underlying
runtimes. Auditing events are likewise considered implementation
specific, but are bound by normal feature compatibility guarantees.
Audit Hook
----------
In order to observe actions taken by the runtime (on behalf of the
caller), an API is required to raise messages from within certain
operations. These operations are typically deep within the Python
runtime or standard library, such as dynamic code compilation, module
imports, DNS resolution, or use of certain modules such as ``ctypes``.
The following new C APIs allow embedders and CPython implementors to
send and receive audit hook messages::
# Add an auditing hook
typedef int (*hook_func)(const char *event, PyObject *args,
void *userData);
int PySys_AddAuditHook(hook_func hook, void *userData);
# Raise an event with all auditing hooks
int PySys_Audit(const char *event, PyObject *args);
The new Python APIs for receiving and raising audit hooks are::
# Add an auditing hook
sys.addaudithook(hook: Callable[[str, tuple]])
# Raise an event with all auditing hooks
sys.audit(str, *args)
Hooks are added by calling ``PySys_AddAuditHook()`` from C at any time,
including before ``Py_Initialize()``, or by calling
``sys.addaudithook()`` from Python code. Hooks cannot be removed or
replaced. For CPython, hooks added from C are global, while hooks added
from Python are only for the current interpreter. Global hooks are
executed before interpreter hooks.
When events of interest are occurring, code can either call
``PySys_Audit()`` from C (while the GIL is held) or ``sys.audit()``. The
string argument is the name of the event, and the tuple contains
arguments. A given event name should have a fixed schema for arguments,
which should be considered a public API (for each x.y version release),
and thus should only change between feature releases with updated
documentation. To minimize overhead and simplify handling in native code
hook implementations, named arguments are not supported.
For maximum compatibility, events using the same name as an event in
the reference interpreter CPython should make every attempt to use
compatible arguments. Including the name or an abbreviation of the
implementation in implementation-specific event names will also help
prevent collisions. For example, a ``pypy.jit_invoked`` event is clearly
distinguished from an ``ipy.jit_invoked`` event. Events raised from
Python modules should include their module or package name in the event
name.
While event names may be arbitrary UTF-8 strings, for consistency across
implementations it is recommended to use valid Python dotted names and
avoid encoding specific details in the name. For example, an ``import``
event with the module name ``spam`` as an argument is preferable to a
``spam module imported`` event with no arguments. Avoid using embedded
null characters or you may upset those who implement hooks using C.
When an event is audited, each hook is called in the order it was added
(as much as is possible), passing the event name and arguments. If any
hook returns with an exception set, later hooks are ignored and *in
general* the Python runtime should terminate - exceptions from hooks are
not intended to be handled or treated as expected occurrences. This
allows hook implementations to decide how to respond to any particular
event. The typical responses will be to log the event, abort the
operation with an exception, or to immediately terminate the process with
an operating system exit call.
When an event is audited but no hooks have been set, the ``audit()``
function should impose minimal overhead. Ideally, each argument is a
reference to existing data rather than a value calculated just for the
auditing call.
As hooks may be Python objects, they need to be freed during
interpreter or runtime finalization. These should not be triggered at
any other time, and should raise an event hook to ensure that any
unexpected calls are observed.
Below in `Suggested Audit Hook Locations`_, we recommend some important
operations that should raise audit events. In general, events should be
raised at the lowest possible level. Given the choice between raising an
event from Python code or native code, raising from native code should be
preferred.
Python implementations should document which operations will raise
audit events, along with the event schema. It is intentional that
``sys.addaudithook(print)`` is a trivial way to display all messages.
Verified Open Hook
------------------
Most operating systems have a mechanism to distinguish between files
that can be executed and those that can not. For example, this may be an
execute bit in the permissions field, a verified hash of the file
contents to detect potential code tampering, or file system path
restrictions. These are an important security mechanism for ensuring
that only code that has been approved for a given environment is
executed.
Most kernels offer ways to restrict or audit binaries loaded and executed
by the kernel. File types owned by Python appear as regular data and
these features do not apply. This open hook allows Python embedders to
integrate with operating system support when launching scripts or
importing Python code.
The new public C API for the verified open hook is::
# Set the handler
typedef PyObject *(*hook_func)(PyObject *path, void *userData)
int PyFile_SetOpenCodeHook(hook_func handler, void *userData)
# Open a file using the handler
PyObject *PyFile_OpenCode(const char *path)
The new public Python API for the verified open hook is::
# Open a file using the handler
io.open_code(path : str) -> io.IOBase
The ``io.open_code()`` function is a drop-in replacement for
``open(abspath(str(pathlike)), 'rb')``. Its default behaviour is to
open a file for raw, binary access. To change the behaviour a new
handler should be set. Handler functions only accept ``str`` arguments.
The C API ``PyFile_OpenCode`` function assumes UTF-8 encoding. Paths
must be absolute, and it is the responsibility of the caller to ensure
the full path is correctly resolved.
A custom handler may be set by calling ``PyFile_SetOpenCodeHook()`` from
C at any time, including before ``Py_Initialize()``. However, if a hook
has already been set then the call will fail. When ``open_code()`` is
called with a hook set, the hook will be passed the path and its return
value will be returned directly. The returned object should be an open
file-like object that supports reading raw bytes. This is explicitly
intended to allow a ``BytesIO`` instance if the open handler has already
read the entire file into memory.
Note that these hooks can import and call the ``_io.open()`` function on
CPython without triggering themselves. They can also use ``_io.BytesIO``
to return a compatible result using an in-memory buffer.
If the hook determines that the file should not be loaded, it should
raise an exception of its choice, as well as performing any other
logging.
All import and execution functionality involving code from a file will
be changed to use ``open_code()`` unconditionally. It is important to
note that calls to ``compile()``, ``exec()`` and ``eval()`` do not go
through this function - an audit hook that includes the code from these
calls is the best opportunity to validate code that is read from the
file. Given the current decoupling between import and execution in
Python, most imported code will go through both ``open_code()`` and the
log hook for ``compile``, and so care should be taken to avoid
repeating verification steps.
File accesses that are not intentionally planning to execute code are
not expected to use this function. This includes loading pickles, XML
or YAML files, where code execution is generally considered malicious
rather than intentional. These operations should provide their own
auditing events, preferably distinguishing between normal functionality
(for example, ``Unpickler.load``) and code execution
(``Unpickler.find_class``).
A few examples: if the file type normally requires an execute bit (on
POSIX) or would warn when marked as having been downloaded from the
internet (on Windows), it should probably use ``open_code()`` rather
than plain ``open()``. Opening ZIP files using the ``ZipFile`` class
should use ``open()``, while opening them via ``zipimport`` should use
``open_code()`` to signal the correct intent. Code that uses the wrong
function for a particular context may bypass the hook, which in CPython
and the standard library should be considered a bug. Using a combination
of ``open_code`` hooks and auditing hooks is necessary to trace all
executed sources in the presence of arbitrary code.
There is no Python API provided for changing the open hook. To modify
import behavior from Python code, use the existing functionality
provided by ``importlib``.
API Availability
----------------
While all the functions added here are considered public and stable API,
the behavior of the functions is implementation specific. Most
descriptions here refer to the CPython implementation, and while other
implementations should provide the functions, there is no requirement
that they behave the same.
For example, ``sys.addaudithook()`` and ``sys.audit()`` should exist but
may do nothing. This allows code to make calls to ``sys.audit()``
without having to test for existence, but it should not assume that its
call will have any effect. (Including existence tests in
security-critical code allows another vector to bypass auditing, so it
is preferable that the function always exist.)
``io.open_code(path)`` should at a minimum always return
``_io.open(path, 'rb')``. Code using the function should make no further
assumptions about what may occur, and implementations other than CPython
are not required to let developers override the behavior of this
function with a hook.
Suggested Audit Hook Locations
==============================
The locations and parameters in calls to ``sys.audit()`` or
``PySys_Audit()`` are to be determined by individual Python
implementations. This is to allow maximum freedom for implementations
to expose the operations that are most relevant to their platform,
and to avoid or ignore potentially expensive or noisy events.
Table 1 acts as both suggestions of operations that should trigger
audit events on all implementations, and examples of event schemas.
Table 2 provides further examples that are not required, but are
likely to be available in CPython.
Refer to the documentation associated with your version of Python to
see which operations provide audit events.
.. csv-table:: Table 1: Suggested Audit Hooks
:header: "API Function", "Event Name", "Arguments", "Rationale"
:widths: 2, 2, 3, 6
``PySys_AddAuditHook``, ``sys.addaudithook``, "", "Detect when new
audit hooks are being added.
"
``PyFile_SetOpenCodeHook``, ``cpython.PyFile_SetOpenCodeHook``, "
", "Detects any attempt to set the ``open_code`` hook.
"
"``compile``, ``exec``, ``eval``, ``PyAst_CompileString``,
``PyAST_obj2mod``", ``compile``, "``(code, filename_or_none)``", "
Detect dynamic code compilation, where ``code`` could be a string or
AST. Note that this will be called for regular imports of source
code, including those that were opened with ``open_code``.
"
"``exec``, ``eval``, ``run_mod``", ``exec``, "``(code_object,)``", "
Detect dynamic execution of code objects. This only occurs for
explicit calls, and is not raised for normal function invocation.
"
``import``, ``import``, "``(module, filename, sys.path,
sys.meta_path, sys.path_hooks)``", "Detect when modules are
imported. This is raised before the module name is resolved to a
file. All arguments other than the module name may be ``None`` if
they are not used or available.
"
"``open``", ``io.open``, "``(path, mode, flags)``", "Detect when a
file is about to be opened. *path* and *mode* are the usual parameters
to ``open`` if available, while *flags* is provided instead of *mode*
in some cases.
"
``PyEval_SetProfile``, ``sys.setprofile``, "", "Detect when code is
injecting trace functions. Because of the implementation, exceptions
raised from the hook will abort the operation, but will not be
raised in Python code. Note that ``threading.setprofile`` eventually
calls this function, so the event will be audited for each thread.
"
``PyEval_SetTrace``, ``sys.settrace``, "", "Detect when code is
injecting trace functions. Because of the implementation, exceptions
raised from the hook will abort the operation, but will not be
raised in Python code. Note that ``threading.settrace`` eventually
calls this function, so the event will be audited for each thread.
"
"``_PyObject_GenericSetAttr``, ``check_set_special_type_attr``,
``object_set_class``, ``func_set_code``, ``func_set_[kw]defaults``","
``object.__setattr__``","``(object, attr, value)``","Detect monkey
patching of types and objects. This event
is raised for the ``__class__`` attribute and any attribute on
``type`` objects.
"
"``_PyObject_GenericSetAttr``",``object.__delattr__``,"``(object,
attr)``","Detect deletion of object attributes. This event is raised
for any attribute on ``type`` objects.
"
"``Unpickler.find_class``",``pickle.find_class``,"``(module_name,
global_name)``","Detect imports and global name lookup when
unpickling.
"
.. csv-table:: Table 2: Potential CPython Audit Hooks
:header: "API Function", "Event Name", "Arguments", "Rationale"
:widths: 2, 2, 3, 6
``_PySys_ClearAuditHooks``, ``sys._clearaudithooks``, "", "Notifies
hooks they are being cleaned up, mainly in case the event is
triggered unexpectedly. This event cannot be aborted.
"
``code_new``, ``code.__new__``, "``(bytecode, filename, name)``", "
Detect dynamic creation of code objects. This only occurs for
direct instantiation, and is not raised for normal compilation.
"
``func_new_impl``, ``function.__new__``, "``(code,)``", "Detect
dynamic creation of function objects. This only occurs for direct
instantiation, and is not raised for normal compilation.
"
"``_ctypes.dlopen``, ``_ctypes.LoadLibrary``", ``ctypes.dlopen``, "
``(module_or_path,)``", "Detect when native modules are used.
"
``_ctypes._FuncPtr``, ``ctypes.dlsym``, "``(lib_object, name)``", "
Collect information about specific symbols retrieved from native
modules.
"
``_ctypes._CData``, ``ctypes.cdata``, "``(ptr_as_int,)``", "Detect
when code is accessing arbitrary memory using ``ctypes``.
"
"``new_mmap_object``",``mmap.__new__``,"``(fileno, map_size, access,
offset)``", "Detects creation of mmap objects. On POSIX, access may
have been calculated from the ``prot`` and ``flags`` arguments.
"
``sys._getframe``, ``sys._getframe``, "``(frame_object,)``", "Detect
when code is accessing frames directly.
"
``sys._current_frames``, ``sys._current_frames``, "", "Detect when
code is accessing frames directly.
"
"``socket.bind``, ``socket.connect``, ``socket.connect_ex``,
``socket.getaddrinfo``, ``socket.getnameinfo``, ``socket.sendmsg``,
``socket.sendto``", ``socket.address``, "``(socket, address,)``", "
Detect access to network resources. The address is unmodified from
the original call.
"
"``member_get``, ``func_get_code``, ``func_get_[kw]defaults``
",``object.__getattr__``,"``(object, attr)``","Detect access to
restricted attributes. This event is raised for any built-in
members that are marked as restricted, and members that may allow
bypassing imports.
"
"``urllib.urlopen``",``urllib.Request``,"``(url, data, headers,
method)``", "Detects URL requests.
"
Performance Impact
==================
The important performance impact is the case where events are being
raised but there are no hooks attached. This is the unavoidable case -
once a developer has added audit hooks they have explicitly chosen to
trade performance for functionality. Performance impact with hooks added
are not of interest here, since this is opt-in functionality.
Analysis using the Python Performance Benchmark Suite [1]_ shows no
significant impact, with the vast majority of benchmarks showing
between 1.05x faster to 1.05x slower.
In our opinion, the performance impact of the set of auditing points
described in this PEP is negligible.
Rejected Ideas
==============
Separate module for audit hooks
-------------------------------
The proposal is to add a new module for audit hooks, hypothetically
``audit``. This would separate the API and implementation from the
``sys`` module, and allow naming the C functions ``PyAudit_AddHook`` and
``PyAudit_Audit`` rather than the current variations.
Any such module would need to be a built-in module that is guaranteed to
always be present. The nature of these hooks is that they must be
callable without condition, as any conditional imports or calls provide
opportunities to intercept and suppress or modify events.
Given it is one of the most core modules, the ``sys`` module is somewhat
protected against module shadowing attacks. Replacing ``sys`` with a
sufficiently functional module that the application can still run is a
much more complicated task than replacing a module with only one
function of interest. An attacker that has the ability to shadow the
``sys`` module is already capable of running arbitrary code from files,
whereas an ``audit`` module could be replaced with a single line in a
``.pth`` file anywhere on the search path::
import sys; sys.modules['audit'] = type('audit', (object,),
{'audit': lambda *a: None, 'addhook': lambda *a: None})
Multiple layers of protection already exist for monkey patching attacks
against either ``sys`` or ``audit``, but assignments or insertions to
``sys.modules`` are not audited.
This idea is rejected because it makes it trivial to suppress all calls
to ``audit``.
Flag in sys.flags to indicate "audited" mode
--------------------------------------------
The proposal is to add a value in ``sys.flags`` to indicate when Python
is running in a "secure" or "audited" mode. This would allow
applications to detect when some features are enabled or when hooks
have been added and modify their behaviour appropriately.
Currently, we are not aware of any legitimate reasons for a program to
behave differently in the presence of audit hooks.
Both application-level APIs ``sys.audit`` and ``io.open_code`` are
always present and functional, regardless of whether the regular
``python`` entry point or some alternative entry point is used. Callers
cannot determine whether any hooks have been added (except by performing
side-channel analysis), nor do they need to. The calls should be fast
enough that callers do not need to avoid them, and the program is
responsible for ensuring that any added hooks are fast enough to not
affect application performance.
The argument that this is "security by obscurity" is valid, but
irrelevant. Security by obscurity is only an issue when there are no
other protective mechanisms; obscurity as the first step in avoiding
attack is strongly recommended (see `this article
<https://danielmiessler.com/study/security-by-obscurity/>`_ for
discussion).
This idea is rejected because there are no appropriate reasons for an
application to change its behaviour based on whether these APIs are in
use.
Why Not A Sandbox
=================
Sandboxing CPython has been attempted many times in the past, and each
past attempt has failed. Fundamentally, the problem is that certain
functionality has to be restricted when executing the sandboxed code,
but otherwise needs to be available for normal operation of Python. For
example, completely removing the ability to compile strings into
bytecode also breaks the ability to import modules from source code, and
if it is not completely removed then there are too many ways to get
access to that functionality indirectly. There is not yet any feasible
way to generically determine whether a given operation is "safe" or not.
Further information and references available at [2]_.
This proposal does not attempt to restrict functionality, but simply
exposes the fact that the functionality is being used. Particularly for
intrusion scenarios, detection is significantly more important than
early prevention (as early prevention will generally drive attackers to
use an alternate, less-detectable, approach). The availability of audit
hooks alone does not change the attack surface of Python in any way, but
they enable defenders to integrate Python into their environment in ways
that are currently not possible.
Since audit hooks have the ability to safely prevent an operation
occurring, this feature does enable the ability to provide some level of
sandboxing. In most cases, however, the intention is to enable logging
rather than creating a sandbox.
Relationship to PEP 551
=======================
This API was originally presented as part of
:pep:`551` Security
Transparency in the Python Runtime.
For simpler review purposes, and due to the broader applicability of
these APIs beyond security, the API design is now presented separately.
:pep:`551` is an informational PEP discussing how to integrate Python into
a secure or audited environment.
References
==========
.. [1] Python Performance Benchmark Suite `<https://github.com/python/performance>`_
.. [2] Python Security model - Sandbox `<https://python-security.readthedocs.io/security.html#sandbox>`_
Copyright
=========
Copyright (c) 2019 by Microsoft Corporation. This material may be
distributed only subject to the terms and conditions set forth in the
Open Publication License, v1.0 or later (the latest version is presently
available at http://www.opencontent.org/openpub/).