Re: [secdir] [Last-Call] Secdir last call review of draft-ietf-netconf-crypto-types-20

[Valery Smyslov <[e-mail]>]

Hi Kent,

 

thank you for addressing my comments. A bit more inline.

 

[removing “last-call” alias]

 

 

Hello Valery.  

 

Thank you for your SecDir review, it is very much appreciated and never too late!

 

More comments below.

 

Thanks,

Kent

 





On Aug 24, 2021, at 2:29 PM, Valery Smyslov via Datatracker <[e-mail]> wrote:

 

Reviewer: Valery Smyslov
Review result: Has Issues

I have reviewed this document as part of the security directorate's ongoing
effort to review all IETF documents being processed by the IESG.  These
comments were written primarily for the benefit of the security area directors.
Document editors and WG chairs should treat these comments just like any other
last call comments.

When I was re-assigned to review this draft the indicated deadline was already
missed by 8 months, so I don't know how relevant the review is. I reviewed
the latest -20 version of the draft instead of -18, that was requested.

 

Perfect.

 





The draft defines common YANG data types and groupings useful for cryptography.
I din't try to check the YANG module itself. 

Issues:

Shouldn't a Privacy Considerations section be added to the draft?
The draft defines quite a lot of privacy-sensitive information (like certificates)
with no restriction on read access (as far as I understand).

 

 

Both the "trust-anchor-cert-grouping” and “end-entity-cert-grouping” groupings have "nacm:default-deny-write” that, to your point, does not restrict reads.  That said, only management protocols having mutual authentication (e.g., SSH and/or TLS based transport) can access the data.  

 

Is your concern that the certificate’s content would be visible to the administrators?  Is your comment on end-entity certificates (containing personally-identifying information), more than trust-anchor-certificates?

 

          Yes, it’s mostly on end-entity certificates, however there may be quite a lot of interesting

          private information besides certificates.

 

          If this information is only visible to the administrators and the used management protocols must 

          have mutual authentication, then it’s probably not a big deal. I would have still added

          a sentence about privacy of the stored data (i.e. that persons, that are allowed to access this data

          are able to learn quite a lot of private information from it). I don’t insist though, it’s up to you.

 

Section 3.5.
While I understand and support the idea, expressed in this section, I think that
the way it is expressed makes it difficult to follow in practice. In general, it's
not always obvious how to estimate the "strength" of the underlying secure transport.
For this reason it's not clear for me how it is supposed to "compare" the 
"strength" of the transport with the "strength" of the keys being transported.

 

I saw language like this once in a DoD setting.  I agree that it is difficult to implement in practice.  I used “SHOULD” (not MUST) to buy some leeway for implementations to be compliant.  Makes sense?

 

          My understanding of using RFC2119 language is that SHOULD is very close to MUST,

          but allows some exceptions. So, I still think that you put a responsibility to make

          security-related decisions on implementers, who often are not experts in this area.

 

FWIW, my YANG-driven server is able to remember what key the client used for authentication (e.g., RSA 2048) and register a callback to test that no greater keys (e.g., 3072 or 4096) are configured by that client.  Additional 

 

          What if the other key configured for the client is X25519? Which is stronger?

 

logic would be needed to prevent a low-strength client from *reading* a high-strength key configured by another client…though the issue can be alternatively resolved by configuring the TLS-stack to prevent low-strength algorithms.

 

          That was my point. I think that it’s better to require (by SHOULD) that only

          those ciphersuites that were “vetted” by IETF (i.e. got “Recommended” status) be used.

          This will make implementers’ life easier.

 

In addition, the requirement, that "Implementations SHOULD fail the write-request if ever
the strength of the private key is greater then the strength of the
underlying transport" looks wrong to me. You don't need to have
1024 bits transport protocol strength to transfer 1024 bit key, since
even for say 256 bits it's infeasible to break.

 

IDK about this.  Again, I saw this constraint once in a DoD setting.  

 

          My (another) point was that there is generally no point to increase

          security strength beyond some level. Currently it is believed

          that 128 bit of symmetric key is infeasible to break (provided the algorithm is not broken itself),

          If you are lucky have full-sized Post Quantum computer, it’ll be 256 bits.

          It’s enough to transfer symmetric keys with say 1024 bits of entropy

          (FWIW). So the requirement that the strength of transport must

          be always greater than the strength of transported key 

          seems not a good requirement to me. Instead require that

          the strength of transport be sufficient to make

          infeasible for an attacker to break it.

          

I think that the better approach would be to advise using strong
ciphersuites for transport protocols defined in corresponding RFCs.
For example, for TLS 1.3 there are ciphersuites marked as "recommended",
that were evaluated by IETF crypto community.

 

I added this sentence:

 

          Implementations SHOULD configure allowed transport 
          algorithms to include only those meeting local

          policy (e.g., listed as "recommended" by the IETF).

 

Good?

 

          Perfect.

 

          Regards,

          Valery.



Section 3.6:
  For instance, AES
  using "EBC" SHOULD NOT be used to encrypt passwords, whereas "CBC"
  mode is okay since it a unique initialization vector (IV) should be
  used for each run.

Not only IV for CBC should be unique, it should be unpredictable (random or pseudo-random).
I also think that lowercase "should" here must be uppercase, or even MUST.

 

Replaced “unique” with “unpredictable”

Replaced “should” with MUST.

 

 

Typos, nits:

Section 3.6: 
  In order to thwart rainbow attacks, algorithms that result in a
  unique output for the same input SHOULD be used. 

s/SHOULD/SHOULD NOT

 

Egads - fixed!

 





  For instance, AES
  using "EBC" SHOULD NOT be used to encrypt passwords, whereas "CBC"
  mode is okay since it a unique initialization vector (IV) should be
  used for each run.

s/EBC/ECB
s/it//
I believe "okay' is a bit of slang, isn't it?

 

Fixed, now reads:

 

          In order to thwart rainbow attacks, algorithms that result
          in a unique output for the same input SHOULD NOT be used.

          For instance, AES using "ECB" SHOULD NOT be used to

          encrypt passwords, whereas "CBC" mode is permissible

          since an unpredictable initialization vector (IV) MUST be

          used for each use.</t>

 

Section 3.8:
  Since the module in this document only define groupings, these
  considerations are primarily for the designers of other modules that
  use these groupings.

s/define/defines

 

Fixed!

 

 

 

Kent, as author