[Openid-specs-ab] Transient Client Secret Extension for OAuth
sakimura at gmail.com
Mon Jul 29 10:46:48 UTC 2013
I was thinking exactly like you do. I am willing to revert my changes.
I kind of understand John's concern when we are using the hash function for
If the attacker is trying to fabricate a new message that has the same hash
value so that the previously calculated signature by the victim still
verifies for the new message. However, in our case here, it does not seem
Let us consider a case that the attacker has a rainbow table such as
lhash1 preimage1 preimage2
There has to be 2^128 rows in the hash table. I do not think it has been
calculated nor is feasible to search through the table within the response
Now, suppose such rainbow table was calculated. Then, using salted hash
would actually constrain the search space to 2^32 rows, which makes it
As long as the rainbow table is not there, it does not seem to matter, but
once it is there, it is safer to be without the salt, I suppose.
As usual, I may be completely wrong but...
2013/7/29 Dirk Balfanz <balfanz at google.com>
> On Mon, Jul 29, 2013 at 12:35 AM, John Bradley <ve7jtb at ve7jtb.com> wrote:
>> Increasing the length of the secret is no help.
> But that's exactly what the salt does: it increases the length of the
> thing that is hashed, thus making it "many orders of magnitude" harder to
> find the preimage. I still don't see why you need a salt here. If you think
> your secret is too small, then create a salt, prepend that to the secret,
> and call the new thing the secret. Salts make sense if the original input
> into your hash function may be in a dictionary of rainbow table. Here the
> client has control over the original input (it generates the secret) and
> can make the probability of that occurring as small as it wants.
>> The thing is that the password is unstructured so any string that
>> computes to the same hash works.
>> If you send a literal that needs to be combined with the password by the
>> server then the hash table has to not only match the hash, the string to
>> produce the hash needs to start with the salt.
>> This make brute forcing the hash may orders of magnitude harder with the
>> same hash size.
> Wait, what? Telling me a prefix of the pre-image will make it _harder_ for
> me to find the pre-image? I don't think so. If anything it'll make it
> easier. What makes it harder is not that I will have to match a certain
> prefix, it's that the search space is bigger with a salt (or with a bigger
> secret, for that matter).
>> On 2013-07-29, at 9:19 AM, Nat Sakimura <sakimura at gmail.com> wrote:
>> Hmmm. I do not actually get it.
>> Salting the hash is equivalent to computing hash ( randome_salt +
>> fixed_secret), is it not?
>> Then, like Brian says, having original secret longer and have it being
>> calculated at run-time has the same effect.
>> i.e., hash (long_dynamic_secret) and hash ( randome_salt + fixed_secret)
>> has the same or better effect as long as
>> len (long_dynmaic_secret ) >= len(randome_salt + fixed_secret).
>> If it is a fixed secret, yes, you are right, but it is not.
>> 2013/7/29 John Bradley <ve7jtb at ve7jtb.com>
>>> If it is just a hash then it can be precomputed. Adding a random salt
>>> that is hashed with the value sent to the token endpoint prevents
>>> pre-computation of the hash.
>>> The issues are the same as with PB-KDF given that we are sending the
>>> hash in the clear.
>>> Yes but if I am a bed person and there is no per password salt I use a
>>> precomputed rainbow table and even if I don't cover the entire space
>>> eventually I will find a collision if I intercept enough messages.
>>> On 2013-07-29, at 8:29 AM, Nat Sakimura <sakimura at gmail.com> wrote:
>>> I have thought of that, and I do not think so.
>>> Adding salt amounts to expanding the entropy of the input string.
>>> So, having enough bit length in the transient secret to start with has
>>> the same effect.
>>> Since the validity period of the transient secret is rather short, you
>>> cannot do the offline attack.
>>> The attacker has to have the rainbow table to start with.
>>> What we want to make sure is that len(tcs) > max_len(available rainbow
>>> 2013/7/29 John Bradley <ve7jtb at ve7jtb.com>
>>>> Thinking about it overnight we need to also have a salt sent with the
>>>> hash, to prevent rainbow tables attacks.
>>>> On 2013-07-28, at 9:39 PM, Nat Sakimura <sakimura at gmail.com> wrote:
>>>> As some of you knows, passing the code securely to a native app on iOS
>>>> platform is next to impossible. Malicious application may register the same
>>>> custom scheme as the victim application and hope to obtain the code, whose
>>>> success rate is rather high.
>>>> We have discussed about it during the OpenID Conenct Meeting at IETF 87
>>>> today, and I have captured the discussion in the form of I-D. It is pretty
>>>> short and hopefully easy to read.
>>>> You can find it at:
>>>> Comments are welcome.
>>>> Nat Sakimura (=nat)
>>>> Chairman, OpenID Foundation
>>>> Openid-specs-ab mailing list
>>>> Openid-specs-ab at lists.openid.net
>>> Nat Sakimura (=nat)
>>> Chairman, OpenID Foundation
>> Nat Sakimura (=nat)
>> Chairman, OpenID Foundation
>> Openid-specs-ab mailing list
>> Openid-specs-ab at lists.openid.net
Nat Sakimura (=nat)
Chairman, OpenID Foundation
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Openid-specs-ab