bug#41456: fix cases where insecure randomness could be used

[Pádraig Brady]

On 22/05/2020 16:47, Taylor Hornby wrote:
> On Fri, May 22, 2020 at 8:17 AM Pádraig Brady <address@hidden> wrote:
> > On 22/05/2020 08:19, Taylor Hornby wrote:
> > > I reported a potential security bug on GitHub:
> > > https://github.com/coreutils/coreutils/pull/32. To save you a click, I'll
> > > copy-paste it here (for context this is on a PR with a fix):
> > >
> > > Comment #1:
> > >
> > > Apologies for submitting on GitHub, it's so much more convenient. I will
> > > understand if no one sees this because I didn't follow the guidelines.
> > >
> > > Justification:
> > >
> > >      - The existing code is dangerous because it can silently fail to seed
> > >      the random number generator securely, either when fopen() fails or when
> > >      read() returns fewer bytes than requested, which can happen if the call
> > >      is interrupted by an interrupt. This is important for utilities like
> > >      shred where cryptographic-quality randomness is important.
> > >      - I removed the bytes_bound stuff because it didn't seem necessary
> > >      anywhere it was used, and if get_nonce is ever called with bytes_bound <
> > >      bufsize, then part of ISAAC's initial state will contain
> > >      timestamps/PIDs, so it will not be uniformly random. Usually, stream
> > >      ciphers like ISAAC require their initial state to be uniformly random,
> > >      otherwise there will be statistical biases in the early output.
> > >
> > > I have not tested all the utilities this affects.
> > >
> > > (Full disclosure is appropriate in this case because any damage has already
> > > been done, fixing the problem in secret would not stop any attacks, but
> > > disclosing might encourage users to stop using the dangerous code and
> > > upgrade.)
> > >
> > > Comment #2:
> > >
> > > This is a more serious issue on Solars, which apparently has a blocking
> > > /dev/random <https://icmconference.org/wp-content/uploads/G11b-Fenwick.pdf>
> > > and NAME_OF_NONCE_DEVICE defaults to /dev/random (see gc-random.m4), or
> > > when NAME_OF_NONCE_DEVICE is overriden to /dev/random with a configure flag
> > > on a Linux system.
> > >
> > > I ran some experiments on a Debian 9 box, and read() from /dev/random
> > > frequently returns very few bytes, sometimes as few as just 6 bytes. This
> > > means, ironically, if someone built the code with /dev/random thinking it
> > > would be more secure, it's actually less secure, because read() will return
> > > fewer bytes and then very little of the ISAAC seed will be random and most
> > > of it will be timestamp/PID/uninitialized memory.
> >
> > Testing on a Solaris 10 box indicates that /dev/random doesn't give short reads.
> > All other systems default to /dev/urandom.
> > coreutils doesn't need cryptographic randomness, so the read from
> > /dev/urandom should be seen as optional, and present to improve randomness when 
> > available.
> > I'm not sure your concerns are justified in the coreutils context.
>
> Thanks for your reply. Yeah, I really doubt it's a problem for most
> use cases. However I do use the shred utility in a context that
> requires high-quality randomness: filling a disk with random data
> prior to using LUKS encryption. Doing so (with good randomness) makes
> it impossible to tell which parts of the disk have ciphertext from
> LUKS and which are "unwritten" (since the randomness pass), preventing
> a small information leak where you can tell how much & where data
> exists on the drive. I suppose instead of using shred, I should just
> do a pass of zeros on the encrypted device, so that it's completely
> filled with ciphertext. I think at the very least a warning should be
> added to shred's help output or manpage that its output cannot be
> relied on to be cryptographically secure.

If you were considering changing from the operations,
one could pass --random-source to shred, which will fail
as you desire if there is not enough random data.

For example you might pass --random-source=/dev/urandom,
though noting that this may result in slower operation
than the internal PRNG.

The default operation is to seed the internal PRNG with 2K
of random data from /dev/urandom. I think you're saying
that's sufficient for your needs, but if there was ever
a short read then the output from coreutils' PRNG would
not be sufficient, and distinguishable from real LUKS encrypted data.
Have you noticed that with an entropy determination utility?
If that was actually the case, then it might be worth
issuing a warning upon short/failed read of the default nonce device,
as it would be both consequential and unlikely.

cheers,
Pádraig