Re: OpenID/Debian PRNG/DNS Cache poisoning advisory [Archief]

Eric Rescorla

08/08/08, 17:35

At Fri, 8 Aug 2008 11:50:59 +0100,
Ben Laurie wrote:
> However, since the CRLs will almost certainly not be checked, this
> means the site will still be vulnerable to attack for the lifetime of
> the certificate (and perhaps beyond, depending on user
> behaviour). Note that shutting down the site DOES NOT prevent the
> attack.
>
> Therefore mitigation falls to other parties.
>
> 1. Browsers must check CRLs by default.

Isn't this a good argument for blacklisting the keys on the client
side?

-Ekr

Dave Korn

08/08/08, 19:35

Eric Rescorla wrote on 08 August 2008 16:06:

> At Fri, 8 Aug 2008 11:50:59 +0100,
> Ben Laurie wrote:
>> However, since the CRLs will almost certainly not be checked, this
>> means the site will still be vulnerable to attack for the lifetime of
>> the certificate (and perhaps beyond, depending on user
>> behaviour). Note that shutting down the site DOES NOT prevent the attack.
>>
>> Therefore mitigation falls to other parties.
>>
>> 1. Browsers must check CRLs by default.
>
> Isn't this a good argument for blacklisting the keys on the client
> side?

Isn't that exactly what "Browsers must check CRLs" means in this context
anyway? What alternative client-side blacklisting mechanism do you suggest?

cheers,
DaveK
--
Can't think of a witty .sigline today....

Eric Rescorla

08/08/08, 19:45

At Fri, 8 Aug 2008 17:31:15 +0100,
Dave Korn wrote:
>
> Eric Rescorla wrote on 08 August 2008 16:06:
>
> > At Fri, 8 Aug 2008 11:50:59 +0100,
> > Ben Laurie wrote:
> >> However, since the CRLs will almost certainly not be checked, this
> >> means the site will still be vulnerable to attack for the lifetime of
> >> the certificate (and perhaps beyond, depending on user
> >> behaviour). Note that shutting down the site DOES NOT prevent the attack.
> >>
> >> Therefore mitigation falls to other parties.
> >>
> >> 1. Browsers must check CRLs by default.
> >
> > Isn't this a good argument for blacklisting the keys on the client
> > side?
>
> Isn't that exactly what "Browsers must check CRLs" means in this context
> anyway? What alternative client-side blacklisting mechanism do you suggest?

It's easy to compute all the public keys that will be generated
by the broken PRNG. The clients could embed that list and refuse
to accept any certificate containing one of them. So, this
is distinct from CRLs in that it doesn't require knowing
which servers have which cert...

-Ekr

Ben Laurie

08/08/08, 19:55

On Fri, Aug 8, 2008 at 5:57 PM, Eric Rescorla <ekr@networkresonance.com> wrote:
> At Fri, 8 Aug 2008 17:31:15 +0100,
> Dave Korn wrote:
>>
>> Eric Rescorla wrote on 08 August 2008 16:06:
>>
>> > At Fri, 8 Aug 2008 11:50:59 +0100,
>> > Ben Laurie wrote:
>> >> However, since the CRLs will almost certainly not be checked, this
>> >> means the site will still be vulnerable to attack for the lifetime of
>> >> the certificate (and perhaps beyond, depending on user
>> >> behaviour). Note that shutting down the site DOES NOT prevent the attack.
>> >>
>> >> Therefore mitigation falls to other parties.
>> >>
>> >> 1. Browsers must check CRLs by default.
>> >
>> > Isn't this a good argument for blacklisting the keys on the client
>> > side?
>>
>> Isn't that exactly what "Browsers must check CRLs" means in this context
>> anyway? What alternative client-side blacklisting mechanism do you suggest?
>
> It's easy to compute all the public keys that will be generated
> by the broken PRNG. The clients could embed that list and refuse
> to accept any certificate containing one of them. So, this
> is distinct from CRLs in that it doesn't require knowing
> which servers have which cert...

It also only fixes this single type of key compromise. Surely it is
time to stop ignoring CRLs before something more serious goes wrong?

Dick Hardt

08/08/08, 20:05

On 8-Aug-08, at 10:11 AM, Ben Laurie wrote:
>
> It also only fixes this single type of key compromise. Surely it is
> time to stop ignoring CRLs before something more serious goes wrong?

Clearly many implementors have chosen to *knowingly* ignore CRLs
despite the security implications, so my take away would be that the
current public key infrastructure is flawed.

-- Dick

Leichter, Jerry

08/08/08, 20:25

On Fri, 8 Aug 2008, Dave Korn wrote:
| > Isn't this a good argument for blacklisting the keys on the client
| > side?
|
| Isn't that exactly what "Browsers must check CRLs" means in this
| context anyway? What alternative client-side blacklisting mechanism
| do you suggest?
Since the list of bad keys is known and fairly short, one could
explicitly check for them in the browser code, without reference to
any external CRL.

Of course, the browser itself may not see the bad key - it may see key
for something that *contains* a bad key. So such a check would not be
complete. Still, it couldn't hurt.

One could put similar checks everywhere that keys are used. Think of it
as the modern version of code that checks for and rejects DES weak and
semi-weak keys. The more code out there that does the check, the faster
bad keys will be driven out of use.

-- Jerry

Dave Korn

08/08/08, 20:25

Eric Rescorla wrote on 08 August 2008 17:58:

> At Fri, 8 Aug 2008 17:31:15 +0100,
> Dave Korn wrote:
>>
>> Eric Rescorla wrote on 08 August 2008 16:06:
>>
>>> At Fri, 8 Aug 2008 11:50:59 +0100,
>>> Ben Laurie wrote:
>>>> However, since the CRLs will almost certainly not be checked, this
>>>> means the site will still be vulnerable to attack for the lifetime of
>>>> the certificate (and perhaps beyond, depending on user
>>>> behaviour). Note that shutting down the site DOES NOT prevent the
>>>> attack.
>>>>
>>>> Therefore mitigation falls to other parties.
>>>>
>>>> 1. Browsers must check CRLs by default.
>>>
>>> Isn't this a good argument for blacklisting the keys on the client
>>> side?
>>
>> Isn't that exactly what "Browsers must check CRLs" means in this
>> context anyway? What alternative client-side blacklisting mechanism do
>> you suggest?
>
> It's easy to compute all the public keys that will be generated
> by the broken PRNG. The clients could embed that list and refuse
> to accept any certificate containing one of them. So, this
> is distinct from CRLs in that it doesn't require knowing
> which servers have which cert...

<scurries off to read CRL format in RFC>

Oh, you can't specify them solely by key, you have to have all the
associated metadata. That's annoying, yes, I understand your point now.

IIRC various of the vendors' sshd updates released in the immediate wake
of the Debian catastrophe do indeed block all the weak keys.

cheers,
DaveK
--
Can't think of a witty .sigline today....

Peter Gutmann

08/08/08, 20:35

Eric Rescorla <ekr@networkresonance.com> writes:

>It's easy to compute all the public keys that will be generated
>by the broken PRNG. The clients could embed that list and refuse
>to accept any certificate containing one of them. So, this
>is distinct from CRLs in that it doesn't require knowing
>which servers have which cert...

You'd also end up with a rather large list for the client to carry around,
which would be especially problematic for lightweight clients. You'd need to
represent it as something like a Bloom filter to avoid this (given that most
users will just click OK on invalid certs, the small false positive rate
shouldn't have much effect either :-).

Peter.

Dan Kaminsky

08/08/08, 20:35

Eric Rescorla wrote:
> At Fri, 8 Aug 2008 17:31:15 +0100,
> Dave Korn wrote:
>
>> Eric Rescorla wrote on 08 August 2008 16:06:
>>
>>
>>> At Fri, 8 Aug 2008 11:50:59 +0100,
>>> Ben Laurie wrote:
>>>
>>>> However, since the CRLs will almost certainly not be checked, this
>>>> means the site will still be vulnerable to attack for the lifetime of
>>>> the certificate (and perhaps beyond, depending on user
>>>> behaviour). Note that shutting down the site DOES NOT prevent the attack.
>>>>
>>>> Therefore mitigation falls to other parties.
>>>>
>>>> 1. Browsers must check CRLs by default.
>>>>
>>> Isn't this a good argument for blacklisting the keys on the client
>>> side?
>>>
>> Isn't that exactly what "Browsers must check CRLs" means in this context
>> anyway? What alternative client-side blacklisting mechanism do you suggest?
>>
>
> It's easy to compute all the public keys that will be generated
> by the broken PRNG. The clients could embed that list and refuse
> to accept any certificate containing one of them. So, this
> is distinct from CRLs in that it doesn't require knowing
> which servers have which cert...
Funnily enough I was just working on this -- and found that we'd end up
adding a couple megabytes to every browser. #DEFINE NONSTARTER. I am
curious about the feasibility of a large bloom filter that fails back to
online checking though. This has side effects but perhaps they can be
made statistically very unlikely, without blowing out the size of a browser.

Updating the filter could then be something we do on a 24 hour
autoupdate basis. Doing either this, or doing revocation checking over
DNS (seriously), is not necessarily a bad idea. We need to do better
than we've been.

Perry E. Metzger

08/08/08, 21:25

"Ben Laurie" <benl@google.com> writes:
>> It's easy to compute all the public keys that will be generated
>> by the broken PRNG. The clients could embed that list and refuse
>> to accept any certificate containing one of them. So, this
>> is distinct from CRLs in that it doesn't require knowing
>> which servers have which cert...
>
> It also only fixes this single type of key compromise. Surely it is
> time to stop ignoring CRLs before something more serious goes wrong?

The problem is, the CRL mechanism itself is also dangerous. Sadly,
clients are required to keep on going if they can't reach a CRL
server. That means that if you DoSing the CRL servers or use DNS
attacks to effectively take them offline, you've also effectively
eliminated the certificate revocation.

I'm not going to tell you that paying attention to CRLs wouldn't be
better than what happens now, but it will not eliminate the
problem. It is too hard to "prove a negative" (that is, to prove to
yourself that no revocation exists.)

The kerberos style of having credentials expire very quickly is one
(somewhat less imperfect) way to deal with such things, but it is far
from perfect and it could not be done for the ad-hoc certificate
system https: depends on -- the infrastructure for refreshing all the
world's certs every eight hours doesn't exist, and if it did imagine
the chaos if it failed for a major CA one fine morning.

One also worries about what will happen in the UI when a certificate
has been revoked. If it just says "this cert has been revoked,
continue anyway?" the wrong thing will almost always happen.

Perry
--
Perry E. Metzger perry@piermont.com

Eric Rescorla

08/08/08, 21:25

At Fri, 08 Aug 2008 10:43:53 -0700,
Dan Kaminsky wrote:
> Eric Rescorla wrote:
> > It's easy to compute all the public keys that will be generated
> > by the broken PRNG. The clients could embed that list and refuse
> > to accept any certificate containing one of them. So, this
> > is distinct from CRLs in that it doesn't require knowing
> > which servers have which cert...
> Funnily enough I was just working on this -- and found that we'd end up
> adding a couple megabytes to every browser. #DEFINE NONSTARTER. I am
> curious about the feasibility of a large bloom filter that fails back to
> online checking though. This has side effects but perhaps they can be
> made statistically very unlikely, without blowing out the size of a browser.

Why do you say a couple of megabytes? 99% of the value would be
1024-bit RSA keys. There are ~32,000 such keys. If you devote an
80-bit hash to each one (which is easily large enough to give you a
vanishingly small false positive probability; you could probably get
away with 64 bits), that's 320KB. Given that the smallest Firefox
build (Windows) is 7.1 MB, this doesn't sound like a nonstarter to me
at all, especially since the browser could download it in the
background.

> Updating the filter could then be something we do on a 24 hour
> autoupdate basis. Doing either this, or doing revocation checking over
> DNS (seriously), is not necessarily a bad idea. We need to do better
> than we've been.

Yes, there are a number of approaches to more efficient CRL
checking, I think that's a separate issue.

-Ekr

Ben Laurie

08/08/08, 21:35

On Fri, Aug 8, 2008 at 7:54 PM, Tim Dierks <tim@dierks.org> wrote:
> Using this Bloom filter calculator:
> http://www.cc.gatech.edu/~manolios/bloom-filters/calculator.html , plus the
> fact that there are 32,768 weak keys for every key type & size, I get
> various sizes of necessary Bloom filter, based on how many key type / sizes
> you want to check and various false positive rates:
> * 3 key types/sizes with 1e-6 false positive rate: 2826759 bits = 353 KB
> * 3 key types/sizes with 1e-9 false positive rate: 4240139 bits = 530 KB
> * 7 key types/sizes with 1e-6 false positive rate: 6595771 bits = 824 KB
> * 7 key types/sizes with 1e-9 false positive rate: 9893657 bits = 1237 KB
>
> I presume that the first 3 & first 7 key type/sizes in this list
> http://metasploit.com/users/hdm/tools/debian-openssl/ are the best to
> incorporate into the filter.
>
> Is there any chance it would be feasible to get a list of all the weak keys
> that were actually certified by browser-installed CAs, or those weak
> certificates? Presumably, this list would be much smaller and would be more
> effectively distributed in Bloom filter form.

Or as a CRL :-)

Nicolas Williams

08/08/08, 21:45

On Fri, Aug 08, 2008 at 02:08:37PM -0400, Perry E. Metzger wrote:
> The kerberos style of having credentials expire very quickly is one
> (somewhat less imperfect) way to deal with such things, but it is far
> from perfect and it could not be done for the ad-hoc certificate
> system https: depends on -- the infrastructure for refreshing all the
> world's certs every eight hours doesn't exist, and if it did imagine
> the chaos if it failed for a major CA one fine morning.

The PKIX moral equivalent of Kerberos V tickets would be OCSP Responses.

I understand most current browsers support OCSP.

> One also worries about what will happen in the UI when a certificate
> has been revoked. If it just says "this cert has been revoked,
> continue anyway?" the wrong thing will almost always happen.

No doubt.

Nicolas Williams

08/08/08, 22:55

On Fri, Aug 08, 2008 at 12:35:43PM -0700, Paul Hoffman wrote:
> At 1:47 PM -0500 8/8/08, Nicolas Williams wrote:
> >On Fri, Aug 08, 2008 at 02:08:37PM -0400, Perry E. Metzger wrote:
> >> The kerberos style of having credentials expire very quickly is one
> >> (somewhat less imperfect) way to deal with such things, but it is far
> >> from perfect and it could not be done for the ad-hoc certificate
> >> system https: depends on -- the infrastructure for refreshing all the
> >> world's certs every eight hours doesn't exist, and if it did imagine
> >> the chaos if it failed for a major CA one fine morning.
> >
> >The PKIX moral equivalent of Kerberos V tickets would be OCSP Responses.
> >
> >I understand most current browsers support OCSP.
>
> ...and only a tiny number of CAs do so.

Not that long ago nothing supported OCSP. If all that's left (ha) is
the CAs then we're in good shape. (OCSP services can be added without
modifying a CA -- just issue the OCSP Responders their certs and let
them use CRLs are their source of revocation information.)

Paul Hoffman

08/08/08, 23:15

At 1:47 PM -0500 8/8/08, Nicolas Williams wrote:
>On Fri, Aug 08, 2008 at 02:08:37PM -0400, Perry E. Metzger wrote:
>> The kerberos style of having credentials expire very quickly is one
>> (somewhat less imperfect) way to deal with such things, but it is far
>> from perfect and it could not be done for the ad-hoc certificate
>> system https: depends on -- the infrastructure for refreshing all the
>> world's certs every eight hours doesn't exist, and if it did imagine
>> the chaos if it failed for a major CA one fine morning.
>
>The PKIX moral equivalent of Kerberos V tickets would be OCSP Responses.
>
>I understand most current browsers support OCSP.

....and only a tiny number of CAs do so.

--Paul Hoffman, Director
--VPN Consortium

Eric Rescorla

08/08/08, 23:25

At Fri, 8 Aug 2008 15:52:07 -0400 (EDT),
Leichter, Jerry wrote:
>
> | > > Funnily enough I was just working on this -- and found that we'd
> | > > end up adding a couple megabytes to every browser. #DEFINE
> | > > NONSTARTER. I am curious about the feasibility of a large bloom
> | > > filter that fails back to online checking though. This has side
> | > > effects but perhaps they can be made statistically very unlikely,
> | > > without blowing out the size of a browser.
> | > Why do you say a couple of megabytes? 99% of the value would be
> | > 1024-bit RSA keys. There are ~32,000 such keys. If you devote an
> | > 80-bit hash to each one (which is easily large enough to give you a
> | > vanishingly small false positive probability; you could probably get
> | > away with 64 bits), that's 320KB. Given that the smallest Firefox
> | > [...]
> You can get by with a lot less than 64 bits. People see problems like
> this and immediately think "birthday paradox", but there is no "birthday
> paradox" here: You aren't look for pairs in an ever-growing set,
> you're looking for matches against a fixed set. If you use 30-bit
> hashes - giving you about a 120KB table - the chance that any given
> key happens to hash to something in the table is one in a billion,
> now and forever. (Of course, if you use a given key repeatedly, and
> it happens to be that 1 in a billion, it will hit every time. So an
> additional table of "known good keys that happen to collide" is worth
> maintaining. Even if you somehow built and maintained that table for
> all the keys across all the systems in the world - how big would it
> get, if only 1 in a billion keys world-wide got entered?)

I don't believe your math is correct here. Or rather, it would
be correct if there was only one bad key.

Remember, there are N bad keys and you're using a b-bit hash,
which has 2^b distinct values. If you put N' entries in the
hash table, the probability that a new key will have the
same digest as one of them is N'/(2^b). If b is sufficiently
large to make collisions rare, then N'=~N and we get
N/(2^b).

To be concrete, we have 2^15 distinct keys, so, the
probability of a false positive becomes (2^15)/(2^b)=2^(b-15).
To get that probability below 1 billion, b+15 >= 30, so
you need about 45 bits. I chose 64 because it seemed to me
that a false positive probability of 2^{-48} or so was better.

-Ekr

Clausen, Martin

12/08/08, 17:25

You could use the SSL Blacklist plugin
(http://codefromthe70s.org/sslblacklist.asp) for Firefox or heise SSL
Guardian
(http://www.heise-online.co.uk/security/Heise-SSL-Guardian--/features/11
1039/) for IE to do this. If presented with a Debian key the show a
warning.

The blacklists are implemented using either a traditional blacklist
(text file) or distributed using DNS.

~martin

-----Original Message-----
From: owner-cryptography@metzdowd.com
[mailto:owner-cryptography@metzdowd.com] On Behalf Of Eric Rescorla
Sent: 8. august 2008 17:06
To: Ben Laurie
Cc: bugtraq@securityfocus.com; security@openid.net; OpenID List;
cryptography@metzdowd.com; full-disclosure@lists.grok.org.uk
Subject: Re: OpenID/Debian PRNG/DNS Cache poisoning advisory

At Fri, 8 Aug 2008 11:50:59 +0100,
Ben Laurie wrote:
> However, since the CRLs will almost certainly not be checked, this
> means the site will still be vulnerable to attack for the lifetime of
> the certificate (and perhaps beyond, depending on user behaviour).
> Note that shutting down the site DOES NOT prevent the attack.
>
> Therefore mitigation falls to other parties.
>
> 1. Browsers must check CRLs by default.

Isn't this a good argument for blacklisting the keys on the client side?

-Ekr

---------------------------------------------------------------------
The Cryptography Mailing List
Unsubscribe by sending "unsubscribe cryptography" to
majordomo@metzdowd.com
---------------------- Deloitte Disclaimer ---------------------
This message (including any attachments) contains confidential
information intended for a specific individual and purpose, and
is protected by law. If you are not the intended recipient, you
should delete this message and are hereby notified that any
disclosure, copying, or distribution of this message, or the
taking of any action based on it, is strictly prohibited.
----------------------------------------------------------------

Ben Laurie

12/08/08, 17:35

On Tue, Aug 12, 2008 at 9:55 AM, Clausen, Martin (DK - Copenhagen)
<mclausen@deloitte.dk> wrote:
> You could use the SSL Blacklist plugin
> (http://codefromthe70s.org/sslblacklist.asp) for Firefox or heise SSL
> Guardian
> (http://www.heise-online.co.uk/security/Heise-SSL-Guardian--/features/11
> 1039/) for IE to do this. If presented with a Debian key the show a
> warning.
>
> The blacklists are implemented using either a traditional blacklist
> (text file) or distributed using DNS.

Browser plugins do not assist RPs.

Ben Laurie

12/08/08, 17:35

Stefan Kanthak

12/08/08, 22:55

Dan Kaminsky wrote:
>
>
> Eric Rescorla wrote:
>> At Fri, 8 Aug 2008 17:31:15 +0100,
>> Dave Korn wrote:
>>
>>> Eric Rescorla wrote on 08 August 2008 16:06:
>>>
>>>
>>>> At Fri, 8 Aug 2008 11:50:59 +0100,
>>>> Ben Laurie wrote:
>>>>
>>>>> However, since the CRLs will almost certainly not be checked, this
>>>>> means the site will still be vulnerable to attack for the lifetime of
>>>>> the certificate (and perhaps beyond, depending on user
>>>>> behaviour). Note that shutting down the site DOES NOT prevent the attack.
>>>>>
>>>>> Therefore mitigation falls to other parties.
>>>>>
>>>>> 1. Browsers must check CRLs by default.
>>>>>
>>>> Isn't this a good argument for blacklisting the keys on the client
>>>> side?
>>>>
>>> Isn't that exactly what "Browsers must check CRLs" means in this context
>>> anyway? What alternative client-side blacklisting mechanism do you suggest?
>>>
>>
>> It's easy to compute all the public keys that will be generated
>> by the broken PRNG. The clients could embed that list and refuse
>> to accept any certificate containing one of them. So, this
>> is distinct from CRLs in that it doesn't require knowing
>> which servers have which cert...
> Funnily enough I was just working on this -- and found that we'd end up
> adding a couple megabytes to every browser.

At least for the weak keys kudos to Debian's OpenSSL maintainer
there exists an extension to Firefox which checks the keys, see <http://codefromthe70s.org/sslblacklist.asp>, as well as c't's
SSLguardian for the Windows Crypto API, see
<http://www.heise-online.co.uk/security/features/111039/0> and
<http://www.heise-online.co.uk/security/features/111039/1>

Stefan

Forrest J. Cavalier III

12/08/08, 23:05

Eric Rescorla wrote:
>
> To be concrete, we have 2^15 distinct keys, so, the
> probability of a false positive becomes (2^15)/(2^b)=2^(b-15).
> To get that probability below 1 billion, b+15 >= 30, so
> you need about 45 bits. I chose 64 because it seemed to me
> that a false positive probability of 2^{-48} or so was better.
>
> -Ekr

Since it's a known set, I think you can use perfect hashing.
There will still be false positives, but presumably no
"bad" keys, nor keys matching the hash everyone agrees on,
are going to be issued after today, right?

Yeah, right.

Leichter, Jerry

12/08/08, 23:15

| > You can get by with a lot less than 64 bits. People see problems
| > like this and immediately think "birthday paradox", but there is no
| > "birthday paradox" here: You aren't look for pairs in an
| > ever-growing set, you're looking for matches against a fixed set.
| > If you use 30-bit hashes - giving you about a 120KB table - the
| > chance that any given key happens to hash to something in the table
| > is one in a billion, now and forever. (Of course, if you use a
| > given key repeatedly, and it happens to be that 1 in a billion, it
| > will hit every time. So an additional table of "known good keys
| > that happen to collide" is worth maintaining. Even if you somehow
| > built and maintained that table for all the keys across all the
| > systems in the world - how big would it get, if only 1 in a billion
| > keys world-wide got entered?)
| I don't believe your math is correct here. Or rather, it would
| be correct if there was only one bad key.
|
| Remember, there are N bad keys and you're using a b-bit hash, which
| has 2^b distinct values. If you put N' entries in the hash table, the
| probability that a new key will have the same digest as one of them is
| N'/(2^b). If b is sufficiently large to make collisions rare, then
| N'=~N and we get N/(2^b).
|
| To be concrete, we have 2^15 distinct keys, so, the probability of a
| false positive becomes (2^15)/(2^b)=2^(b-15). To get that probability
| below 1 billion, b+15 >= 30, so you need about 45 bits. I chose 64
| because it seemed to me that a false positive probability of 2^{-48}
| or so was better.
You're right, of course - I considered 32,000 to be "vanishingly small"
compared to the number of hash values, but of course it isn't. The
perils of looking at one number just as decimal and the other just in
exponential form....

In any case, I think it's clear that even for extremely conservative
"false hit" ratios, the table size is quite reasonable. You wouldn't
want the table on your smart card or RFID chip, perhaps, but there even
a low-end "smartphone" would have no problems.

-- Jerry

Tim Dierks

12/08/08, 23:35

[Sorry for duplicates, but I got multiple requests for a non-HTML
version, and I didn't want to fork the thread. Also sorry for
initially sending HTML; I didn't realize it was so abhorrent these
days. ]

On Fri, Aug 8, 2008 at 1:43 PM, Dan Kaminsky <dan@doxpara.com> wrote:
>>
>> It's easy to compute all the public keys that will be generated
>> by the broken PRNG. The clients could embed that list and refuse
>> to accept any certificate containing one of them. So, this
>> is distinct from CRLs in that it doesn't require knowing which servers have which cert...
>
> Funnily enough I was just working on this -- and found that we'd end up adding a couple megabytes to every browser. #DEFINE NONSTARTER. I am curious about the feasibility of a large bloom filter that fails back to online checking though. This has side effects but perhaps they can be made statistically very unlikely, without blowing out the size of a browser.

Using this Bloom filter calculator:
http://www.cc.gatech.edu/~manolios/bloom-filters/calculator.html ,
plus the fact that there are 32,768 weak keys for every key type &
size, I get various sizes of necessary Bloom filter, based on how many
key type / sizes you want to check and various false positive rates:
* 3 key types/sizes with 1e-6 false positive rate: 2826759 bits = 353 KB
* 3 key types/sizes with 1e-9 false positive rate: 4240139 bits = 530 KB
* 7 key types/sizes with 1e-6 false positive rate: 6595771 bits = 824 KB
* 7 key types/sizes with 1e-9 false positive rate: 9893657 bits = 1237 KB

I presume that the first 3 & first 7 key type/sizes in this list
http://metasploit.com/users/hdm/tools/debian-openssl/ are the best to
incorporate into the filter.

Is there any chance it would be feasible to get a list of all the weak
keys that were actually certified by browser-installed CAs, or those
weak certificates? Presumably, this list would be much smaller and
would be more effectively distributed in Bloom filter form.

- Tim

Florian Weimer

13/08/08, 00:05

* Eric Rescorla:

> Why do you say a couple of megabytes? 99% of the value would be
> 1024-bit RSA keys. There are ~32,000 such keys.

There are three sets of keys, for big-endian 32-bit, little-endian
32-bit and little-endian 64-bit. On top of that, "openssl genrsa"
generates different keys depending on the existence of $HOME/.rnd (and
-3 creates yet another set of keys, but this is more in the league of
"different key length"). If the library is used for key generation
(instead of the command line tool), different keys might result.

On the other hand, the on-disk size would be comparable to the phishing
filter database.

Part of the problem of the CRL approach is that CAs usually have
policies against obtaining private keys and therefore can't prove to the
customer that their keys are compromised. And adding a CRL entry when
the customer isn't convinced that they've got a problem is probably not
a good idea, either.