Here's an example in Sage, if the verbosity of OpenSSL code annoys you as much as it does me:
n = 115792089210356248762697446949407573529996955224135760342422259061068512044369
p = 2^256 - 2^224 + 2^192 + 2^96 - 1
a = -3
b = 41058363725152142129326129780047268409114441015993725554835256314039467401291
E = EllipticCurve(GF(p), [a,b])
P = E(48439561293906451759052585252797914202762949526041747995844080717082404635286,
36134250956749795798585127919587881956611106672985015071877198253568414405109)
"""
Q = E(91120319633256209954638481795610364441930342474826146651283703640232629993874,
80764272623998874743522585409326200078679332703816718187804498579075161456710)
"""
e = 0x12345678
ei = inverse_mod(e, n)
Q = e*P
def to_bin(x): return ("%060x" % x).decode('hex')
def from_bin(x): return int(x.encode('hex'), 16)
def dual_ec_drbg(s, len):
out = ''
for i in range(0, len, 30):
x, y = (s*P).xy()
s = Integer(x)
x, y = (s*Q).xy()
r = Integer(x) % 2**240
out += to_bin(r)
return out[0:len]
def recover_s(s0, s1):
for i in range(2**16): # For all possible 256-bit x
r = i*2**240 + s0 #
if E.is_x_coord(r): # is valid x?
R = E.lift_x(r) # Lift it to a point
x, y = (R*ei).xy() # Undo s*Q
s = Integer(x)
x, y = (s*Q).xy() # Check if it matches next observed block
if (Integer(x) % 2**240) == s1:
return s # done
return None
import os
# Get 3 blocks: 2 are for the break, the other to confirm prediction
s = dual_ec_drbg(from_bin(os.urandom(16)), 30*3)
s0 = from_bin(s[ 0:30])
s1 = from_bin(s[30:60])
s2 = from_bin(s[60:90])
# Recover internal state
s = recover_s(s0, s1)
# Now try predicting something
t = dual_ec_drbg(s, 30)
assert( from_bin(t) == s2 )
print "Done"
"I did not break the official algorithm. I do not know the secret value used to compute the Q constant, and thus cannot break the default implementation. Only NSA (and people with access to the key) can exploit the PRNG weakness."
I found this note interesting. How big is the secret value used to compute the Q constant? Is it a single static value or does it vary? Would it be possible to brute force this? I'm not a crypto expert and want to understand this a bit better.
One of the big arguments about the NSA introducing weaknesses into these algorithms is "this makes them insecure for everyone and flaws exploited by the government could be exploited by anyone", but this makes it sound like ONLY the NSA could exploit this.
I'm not saying this is better. I just think, if true, it's an interesting discussion point in the debate.
> but this makes it sound like ONLY the NSA could exploit this.
Exactly right, and that's why NSA can still claim with a straight face that they were not trying to go around breaking crypto in a generic fashion. Like Clipper and Lotus Notes's initial crypto support, NSA was trying to make the crypto safe against everyone but NSA.
Now this is one area where I disagree totally with NSA's actions, but it is true that the way they went about weakening Dual EC DRGB is more nuanced than simply inserting a secret backdoor that anyone could exploit.
Right. I think the big issue is that we are one NSA key leak away from breaking every product that uses this type of encryption. It introduces a single point of failure.
TLS provides forward secrecy though, so leaking the SSL keys doesn't mean anything for those exabytes of recorded transactions the NSA is presumably saving for a rainy day. If the secret keys for the default Dual_EC_DRBG configuration are leaked, aren't there more serious implications?
Dual_EC_DRBG is a pseudo random number generator, so if it's compromised you could use the internal state to recreate the sequence of pseudo random digits from that point forward. SSL/TLS is protocol which incorporates a suite of algorithms for authentication, key exchange and encryption. So on that level, like the sibling comment mentions, they aren't really comparable. The part of the SSL/TLS protocol that really depends on the certificate is the authentication, though - Diffie Helmann can be done in the clear without compromising its security, and once you've established the key any algorithm can be used to encrypt. When you revoke a certificate, you're essentially saying that you suspect someone else may be able impersonate you on the internet, so don't trust anyone authenticating with it. The algorithm is secure, but the certificate is not.
The point I was trying to make is that if you suspect someone has compromised your certificate, you can revoke it to reestablish secure communications. If you suspect your random number generator has been compromised, you can likewise just change the configuration (assuming it's not hardcoded into some piece of cryptographic equipment).
I'm confused how this is valid, since he seems to be using the OpenSSL code without the patch[1] that actually makes Dual EC work and his patch doesn't (to my C-ignorant eyes) include the fix either. Does it fix it in another way?
perhaps it is related to this point, since that bug causes the algorithm to get "stuck".
Why wasn't this bug caught in the FIPS 140-2 validation testing?
- ---------------------------------------------------------------
Not only the original validation (#1747) but many subsequent validations
and platforms have successfully passed the CAVP[5] algorithm tests ...
several hundred times now. That's a lot of fail.
In test mode the implementation works fine both with and without
additional data. In free running mode the bug is triggered by additional
data on the first call, which is done automatically by the "FIPS capable"
OpenSSL.
Frankly the FIPS 140-2 validation testing isn't very useful for catching
"real world" problems.
No, they plan on removing the PRNG from their release. I understand better why my POC worked without triggering the bug. From the mail archive:
". When the discard occurs the data must not be output and the Dual EC DRBG state must be updated, but that state update isn't done. In the case of no additional input this has no effect, but additional input is used by the "FIPS capable" OpenSSL. Note that additional input does not effectively defeat the backdoor vulnerability[3]."
I do not use the reseed functionality either, because I only generate two or three output blocs and never call an explicit reseed.
"An" can also be used if the first letter is an h and the first syllable is unstressed (an historical event is the example on the Wikipedia page), but, I'm pretty sure the h in hashing is in a stressed syllable, so that doesn't count either.
Reminds me of some native English speakers who say "an historic ..." Pretty common even in accents that won't drop the H. Seems like that word is an outlier though.