[Cryptography] AES [was NSA and cryptanalysis]

[Dave Howe]

On 16/09/2013 23:39, Perry E. Metzger wrote:
> On Mon, 16 Sep 2013 11:54:13 -1000 Tim Newsham
> <tim.newsham at gmail.com> wrote:
>> - A backdoor that leaks cryptographic secrets
>>
>> consider for example applications using an intel chip with
>> hardware-assist for AES. You're feeding your AES keys
>> directly into the cpu. Any attacker controlling the cpu has
>> direct access and doesn't have to do any fancy pattern matching
>> to discover the keys. Now if that CPU had a way to export
>> some or all of the bits through some channel that would also
>> be passively observable, the attacker could pull off an offline
>> passive attack.
>>
>> What about RNG output? What if some bits were redundantly
>> encoded in some of the RNG output bits which where then
>> used directly for tcp initial sequence numbers?
>>
>> Such a backdoor would be feasible.
> It might be feasible in theory (and see the Illinois Malicious
> Processor as an example) but I think it would be hard to pull off
> well -- too hard to account for changes in future code, too hard to
> avoid detection of what you've done.
Not sure this is true. If instead of leaking via the RNG, you leak via
the cryptographic libraries *and* the windows socket libraries, then
while there are probably two different teams involved, there is only one
manufacturer - Microsoft.

Ok that would exclude non-windows systems, which in this world of BYOD
means an increasing number of ios or android devices - but the odds of
one end or the other of any given exchange being a MS platform are
good.  Provided the cryptographic libraries are queried in a specific
manner for tcp sequence numbers (which can be enforced) the winsock team
never need know how those are generated, leaving just the cryptographic
library holding both the input and output.



>
> On the other hand, we know from the press reports that several
> hardware crypto accelerators have been either backdoored or
> exploited. In those, leaking key material to observers in things like
> IVs or choices of nonces might be quite feasible. Such devices are
> built to be tamper resistant so no one will even notice if you add
> features to try to conceal the "extra functionality" of the device.
>
> For the Intel chips, I suspect that if they've been gimmicked, it
> will be more subtle, like a skew in the RNG that could be explained
> away as a manufacturing or design error. That said, things like the
> IMP do give one pause. And *that* said, if you're willing to go as
> far as what the IMP does, you no longer need to simply try to leak
> information via the RNG or other crypto hardware, you can do far far
> worse.
>
> (For those not familiar with the Illinois Malicious Processor:
> https://www.usenix.org/legacy/event/leet08/tech/full_papers/king/king_html/
> )
>
> Perry