Gonzalo,
Most of this looks good; I have a few comments:
(1) In the two places where there are general recommendations that
are not specific to BFCP (IP address selection and use of
SubjectAltName in certs in preference to Subject), it would
be good to note that these are general recommendations
that are not specific to BFCP, **but** please consult your
AD on whether and how to do this, as these are cross-area
issues in the IETF. The existing text is ok.
(2) I would change the sentence about PBKDF2 key strengthening,
as the use of "security" is questionable:
OLD:
Because such key derivation functions may incorporate
iteration functions for key strengthening they provide some
additional security against dictionary attacks.
NEW:
Because such key derivation functions may incorporate
iteration functions for key strengthening they provide some
additional protection against dictionary attacks by
increasing the amount of work that the attacker must perform.
(3) Regarding the "obtain" language, I would make the following
change to explain why the attacks don't apply -
OLD:
When the keys are randomly generated and of sufficient length,
these attacks do not obtain.
NEW:
When the keys are randomly generated and of sufficient length,
dictionary attacks are not effective because such keys are
highly unlikely to be in the attacker's dictionary.
Thanks,
--David
----------------------------------------------------
David L. Black, Senior Technologist
EMC Corporation, 176 South St., Hopkinton, MA 01748
+1 (508) 293-7953 FAX: +1 (508) 293-7786
black_david@xxxxxxx Mobile: +1 (978) 394-7754
----------------------------------------------------
-----Original Message-----
From: Gonzalo Camarillo [mailto:Gonzalo.Camarillo@xxxxxxxxxxxx]
Sent: Sunday, June 17, 2007 5:52 AM
To: Black, David
Cc: ietf@xxxxxxxx; xcon@xxxxxxxx; tsv-dir@xxxxxxxx; Cullen Jennings
Subject: Re: tsv-dir review of draft-ietf-xcon-bfcp-connection-04
Hi David,
thanks for your comments. Answers inline:
Black_David@xxxxxxx wrote:
I've reviewed this document as part of the transport area
directorate's ongoing effort to review key IETF documents.
These comments were written primarily for the transport
area directors, but are copied to the document's authors
for their information and to allow them to address any
issues raised.
This is a relatively straightforward draft about how a
client opens a TCP connection to a BFCP server when it
has the server's transport address information.
Section 3 ventures into the area of IP address selection -
it references RFC 3484 (which is good) and then goes on to
make additional comments and recommendations on usage and
state of deployment of the techniques specified in RFC 3484.
While there's nothing technically wrong with this text, the
additional comments and recommendations are not specific
to BFCP, and may belong in a more generic document.
Given that such a generic document does not exist, we need to
give these
recommendations here.
Section 4 starts with "All BFCP entities implement TLS ..."
That is correct, as RFC 4582 requires this, but it would be
better to cite RFC 4582 as part of this statement, e.g.,
"[RFC 4582] requires that all BFCP entities implement TLS ..."
What about performing the following change?
OLD:
All BFCP entities implement TLS [7] and SHOULD use it in all their
connections.
NEW:
[RFC 4582] requires that all BFCP entities implement TLS and
recommends
that they use it in all their connections.
In the second paragraph of Section 4, I would change
"can request the use of TLS" to "SHOULD request the use
of TLS".
OK, I will fix it.
Section 5.1 specifies that SubjectAltName identities in
certificates are to be preferred to Subject identities.
Is this specific to BFCP or more general?
We got that recommendation from our security adviser. I do not know
whether this will be documented in a generic document at some point.
The following text appears to be an oversight:
If the client knows the server's IP address, the iPAddress
subjectAltName must be present in the certificate and must
exactly match the IP address known to the client.
The client *always* knows the server's IP address (e.g.,
DNS returns it). I think the intended case here is that
the client contacts the server using the server's IP address
and as a result does not know the server's hostname. Rephrasing
in that sort of fashion would also express a preference for
hostname as certificate identity, which I believe is desirable.
What about performing the following change?:
OLD:
If the client knows the server's IP address, the iPAddress
subjectAltName must be present in the certificate and must exactly
match the IP address known to the client.
NEW:
If the client does not know the server's hostname and contacts the
server directly using the server's IP address, the iPAddress
subjectAltName must be present in the certificate and must exactly
match the IP address known to the client.
Section 6 disturbingly shifts between "password" and
"pre-shared key" and appears to get a few things wrong in
the process. To begin with, the statement that "TLS PSK mode
is subject to offline dictionary attacks." is false when
the PSK is high-entropy. OTOH, it is correct when the PSK
is low-entropy (e.g., a password, or derived from a password
without introduction of additional entropy). The discussion
in Section 7.2 of RFC 4279 applies, especially the last
paragraph about PSK generation. The section needs to be
carefully revised to distinguish between "password" and
"pre-shared key", especially given the mention of use of
PBKDF2 to generate the latter from the former.
what about performing the following change?:
OLD:
TLS PSK mode is subject to offline dictionary attacks. In DHE and
RSA modes, an attacker who can mount a single man-in-the-middle
attack on a client/server pair can then mount a dictionary attack
on
the password. In modes without DHE or RSA, an attacker who can
record communications between a client/server pair can mount a
dictionary attack on the password. Accordingly, it is RECOMMENDED
that where possible clients use certificate-based server
authentication ciphersuites with PSK in order to defend against
dictionary attacks.
In addition, passwords SHOULD be chosen with enough entropy to
provide some protection against dictionary attacks. Because the
entropy of text varies dramatically and is generally far less than
that of an equivalent random bitstring, no hard and fast rules
about
password length are possible. However, in general passwords
SHOULD
be chosen to be at least 8 characters and selected from a pool
containing both upper and lower case, numbers, and special
keyboard
characters (note that an 8-character ASCII password has a maximum
entropy of 56 bits and in general far lower). FIPS PUB 112 [11]
provides some guidance on the relevant issues. If possible,
passphrases are preferable to passwords. In addition, a
cooperating
client and server pair MAY choose to derive the TLS PSK shared key
from the passphrase via a password-based key derivation function
such
as PBKDF2 [2].
NEW:
TLS PSK simply relies on a pre-shared key without specifying the
nature of the key. In practice, such keys have two sources: text
passwords and randomly generated binary keys. When keys are
derived
from passwords, TLS PSK mode is subject to offline dictionary
attacks. In DHE and RSA modes, an attacker who can mount a single
man-in-the-middle attack on a client/server pair can then mount a
dictionary attack on the password. In modes without DHE or RSA, an
attacker who can record communications between a client/server
pair
can mount a dictionary attack on the password. Accordingly, it is
RECOMMENDED that where possible clients use certificate-based
server authentication ciphersuites with password-derived PSKs,
in order to defend against dictionary attacks.
In addition, passwords SHOULD be chosen with enough entropy to
provide some protection against dictionary attacks. Because the
entropy of text varies dramatically and is generally far less than
that of an equivalent random bitstring, no hard and fast rules
about
password length are possible. However, in general passwords
SHOULD
be chosen to be at least 8 characters and selected from a pool
containing both upper and lower case, numbers, and special
keyboard
characters (note that an 8-character ASCII password has a maximum
entropy of 56 bits and in general far lower). FIPS PUB 112 [11]
provides some guidance on the relevant issues. If possible,
passphrases are preferable to passwords. In addition, a
cooperating
client and server pair MAY choose to derive the TLS PSK shared key
from the passphrase via a password-based key derivation function
such
as PBKDF2 [2]. Because such key derivation functions may
incorporate
iteration functions for key strengthening they provide some
additional security against dictionary attacks.
When the keys are randomly generated and of sufficient length,
these attacks do not obtain. Where possible, keys SHOULD be
generated using a strong random number generator as specified
in RFC 4086 [REF]. A minimum key length of 80 bits SHOULD be used.
Thanks,
Gonzalo