Re: [TLS] [re-send] draft-ietf-tls-exported-authenticator IESG review

[Jonathan Hoyland <[e-mail]>]

Hi Nick, all,

So IIUC the EAP-TLS issue refers to the problem where EAP-TLS clients
couldn't determine whether they were authenticated at the end of the TLS
1.3 handshake because the server can reject the client certificate but
continue on with the handshake.

This is the expected / RFC behaviour, per E.1.2:

 A client that has sent authentication data to a server, either during

 the handshake or in post-handshake authentication, cannot be sure

 whether the server afterwards considers the client to be

 authenticated or not. If the client needs to determine if the server

 considers the connection to be unilaterally or mutually

 authenticated, this has to be provisioned by the application layer.

IIRC the solution was to (ab)use the close_notify alert to signal to the
client that the server had accepted its authentication.

With respect to EAs there are two issues.

   1.

   Until the server's first application data message arrives at the client
   the client doesn't know whether some adversary tampered with its final
   flight.
   2.

   When a client sends an EA it doesn't know whether the server accepted
   its handshake auth (whether or not the server has sent an application data
   message.)

Message Tampering:

Ben's proposal to create a new exporter key that includes the client auth
messages ensures agreement on the messages, but delays the point at which
client auth can be sent until after the server's first application message.

Given that the EA protocol can be run out-of-band, and there is no
guarantee that any application messages will be sent, this may be useful.
Although assuming that an application level system is used to confirm auth
I can't actually think of a way this potential discrepancy might be
leveraged.

If a server sent an CertificateRequest in its handshake and also sent
AuthenticatorRequest out-of-band after its Finished message a malicious
client might withhold its Finished message, until after it had responded to
the EA. It is difficult to imagine a scenario where this would lead to a
bad outcome, but if this was part of the threat model the server can simply
not request EAs until after it has received the client’s Finished message.

Auth Confirmation:

Ben's suggestion that we change the key derivation doesn't affect the
client's auth status, because the server may agree on the cert bytes, but
have silently rejected the cert. This still requires application level
confirmation, so this doesn't directly have any bearing on fixing the
EAP-TLS style issue.

Joint Auth:

One potentially interesting property that appears to drop out of Ben's
construction is that EAs sent by the client are bound to the client's
handshake auth. With application layer mechanisms to confirm auth status,
the client can use an EA to prove that the handshake cert was used
honestly, even if the handshake cert has been compromised. I'm not sure
that it's super useful, but it's def. Interesting.

Overall, I would say there are 2 sensible paths forward.

   1.

   Add a section to the “Security Considerations” section along the lines
   of: “In TLS 1.3 the client’s final flight is not agreed until the first
   application message. Because EAs can be negotiated out-of-band it is
   possible to agree EAs without agreeing on the entire transcript. Servers¹
   SHOULD send application data before sending an AuthenticatorRequest to
   the client. If there is no application data to send the server MAY send a
   NewSessionTicket²."

   Clients do not need to do this, because the server will not accept any
   application data messages before it receives the client’s Finished
   message.
   2.

   Move to the new exporter key variant as Ben suggested. This makes EAs
   contingent on the client’s final flight. It provides some level of
   tamper-resistance, at a cost of an extra half RTT. This is also more robust
   when applied to things like DTLS where application layer messages can get
   lost.

We can do either of these. My first choice would be to just do 1, but I’d
also be happy with 2. I've made a PR for 1 here
<https://github.com/tlswg/tls-exported-authenticator/pull/76>.

Regards,

Jonathan

¹ Technically this is only necessary if the server sent a CertificateRequest
in the handshake, but a blanket SHOULD seems simpler to me. If a server
never sends CertificateRequests maybe this is overly expensive.

² An NST is bound to the resumption_master_secret and thus to the client
auth bytes.

On Tue, 24 Aug 2021 at 00:29, Nick Sullivan <nick=
[e-mail]> wrote:

> Hello TLSWG and IESG reviewers,
>
> This is a compendium of responses to the various reviews of the
> document. There are a few remaining open questions to address that I
> hope we can resolve in this thread.
>
> I’ve compiled the changes to the document in response to the comments in
> Github:
> https://github.com/tlswg/tls-exported-authenticator/pull/74
>
> Responses welcome!
>
> =====
>
> Open questions for debate
>
> Should we define an optional API to generate the
> certificate_request_context?
> [ns] Proposed solution: no, let the application decide
>
> Should we remove the SHOULDs in Section 7, as they have little to do
> with interoperability?
> [ns] Proposed solution: leave the text as is.
>
> Are there any security issues caused by the fact that the Exported
> Authenticator is based on the Exporter Secret, which does not
> incorporate the entire transcript:
>
>   Using an exporter value as the handshake context means that any client
>   authentication from the initial handshake is not cryptographically
>   digested into the exporter value.  In light of what we ran into with
>   EAP-TLS1.3, do we want to revisit whether we're still happy with that?
>   We should in practice still benefit from the implicit confirmation that
>   anything the server participates in after receiving the Client Finished
>   means the server properly validated it and did not abort the handshake,
>   but it's a bit awkward, especially since the RFC 8446 post-handshake
>   authentication does digest the entire initial handshake transcript.
>   Would we feel any safer if an exporter variant was available that
>   digested the entire initial handshake transcript and we could use that?
>
> [ns] This is a very good question. I didn’t follow the EAP-TLS1.3
> issue closely, but this does seem to imply an imperfect match with
> post-handshake authentication in the case of mutually authenticated
> connections. I would like to suggest that Jonathan Hoyland (cc'd), who
> did the formal analysis on the protocol, provide his input.
>
> A comment from Yaron Scheffer/Benjamin Kaduk on the IANA considerations:
>
> My understanding is that the intent of this document is to only allow
> "server_name" to appear in the ClientCertificateRequest structure that
> otherwise uses the "CR" notation in the TLS 1.3 column of the TLS
> ExtensionType Values registry, but not to allow for "server_name" in
> authenticator CertificateRequests or handshake CertificateRequests.
> Assuming that's correct, the easiest way to indicate that would be if
> there was a "Comment" column in the registry that could say "only used
> in ClientCertificateRequest certificate requests and no other
> certificate requests", but there is not such a column in the registry.
> I think it could also work to be much more clear in the IANA
> considerations of this document as to why the "CR" is being added and
> what restrictions there are on its use, even if that's not quite as
> clean of a solution.
>
> [ns] Proposed solution:
> Add text to section 8.1 to clarify why the CR is being added and what
> restrictions there are to its use, as is done in
> https://github.com/tlswg/tls-exported-authenticator/pull/74.
>  Alternate solution:
> Propose a new column, as done in
> https://github.com/tlswg/tls-exported-authenticator/pull/75
>
> =====
>
> Comments and responses
>
> Proposals to address IESG comments below are here:
> https://github.com/tlswg/tls-exported-authenticator/pull/74
>
> Martin Duke (No Objection):
>
> - I come away from this not being entirely sure if this document applies to
> DTLS or not. There is the one reference in the intro to DTLS, but it's not
> in
> the abstract, nor anywhere else. Assuming that the Sec. 1 reference is not
> some
> sort of artifact, the document would benefit from a liberal sprinkling of
> 's/TLS/(D)TLS' (but this would not apply to every instance)
>
> [ns] DTLS does not explicitly define a post-handshake authentication,
> so not all references made sense to update, but I updated all the
> relevant references.
>
> - If (D)TLS 1.2 is REQUIRED to implement, then does this document not
> update
> those RFCs?
>
> [ns] It doesn’t change existing versions of (D)TLS and it is not
> mandatory to include this functionality, so I don’t think that it
> updates them. It’s a separate and optional feature built on top of
> TLS. In any case, the text is changed to make it explicit that this is
> a minimum requirement.
>
> NITS:
>
> - Sec 1. I think you mean Sec 4.2.6 of RFC 8446, not 4.6.3.
> [ns] It’s actually neither, it’s 4.6.2. Sec. 4.2.6 defines the
> extension that the client uses to advertise support for post-handshake
> authentication.
>
> - Sec 4 and 5. "use a secure with..." ?
> [ns] Added “transport”
>
> - Sec 4. s/messages structures/message structures
> [ns] Fixed
>
>
>
>
> Erik Kline's No Objection
>
> ----------------------------------------------------------------------
> COMMENT:
> ----------------------------------------------------------------------
>
> [ sections 4 & 5 ]
>
> * "SHOULD use a secure with" ->
>   "SHOULD use a secure communications channel with" or
>   "SHOULD use a secure transport with" or something?
> [ns] Fixed in previous change
>
> * "as its as its" -> "as its"
> [ns] Fixed
>
> [ section 5.2.3 ]
>
> * I think the first sentence of the first paragraph may not be a
>   grammatically complete sentence?
> [ns] Fixed
>
> [ section 7 ]
>
> * "following sections describes APIs" ->
>   "following sections describe APIs"
> [ns] Fixed
>
>
> Francesca Palombini's No Objection
> Thank you for the work on this document, and thank you to the doc shepherd
> for
> the in-depth background. Please find some comments below.
>
> Francesca
>
> 1. -----
>
>    TLS (or DTLS) version 1.2 [RFC5246] [RFC6347]. or later are REQUIRED
>    to implement the mechanisms described in this document.
>
> FP: Does this mechanism applies to DTLS as well? The sentence above, the
> normative reference to DTLS 1.2, and the IANA section 8.2 would imply so.
> However, this is the only time DTLS is mentioned in the body of the
> document,
> excluding IANA considerations. I would like it to be made explicit that
> this
> mechanism can be used for DTLS (if that's the case) both in the abstract
> and in
> the introduction, add any useful considerations about using this mechanism
> with
> DTLS vs TLS, and possibly add a sentence in the introduction stating that
> DTLS
> 1.X uses what specified for TLS 1.X. Two examples of why that would be
> useful
> are the following sentences:
>
>    using an exporter as described in Section 4 of [RFC5705] (for TLS
>    1.2) or Section 7.5 of [TLS13] (for TLS 1.3).  For TLS 1.3, the
>
>    "signature_algorithms" and "signature_algorithms_cert" extension, as
>    described in Section 4.2.3 of [TLS13] for TLS 1.3 or the
>    "signature_algorithms" extension from Sections 7.4.2 and 7.4.6 of
>    [RFC5246] for TLS 1.2.
>
> which makes it clear what applies for TLS but not for DTLS.
>
> [ns] Fixed in previous change
>
>
> 2. -----
>
>    used to send the authenticator request SHOULD use a secure with
>    equivalent security to TLS, such as QUIC [QUIC-TLS], as its as its
>
> FP: What are the implications of not using such a secure transport
> protocol?
> Why is it just RECOMMENDED and not MANDATED? nits: missing word "use a
> secure
> with" ; remove one of the duplicated "as its". (Note: this text appears
> again
> with the same typos for the authenticator in section 5)
>
> [ns] There are no known security implications of using an insecure
> transport here, just privacy implications. This point was debated on
> list (see Yaron Sheffer’s comments) and it was agreed that because
> there may be alternate deployment scenarios that involve transport
> outside the original tunnel, TLS-equivalent security is not a mandated
> requirement. Typos fixed in previous change.
>
>
> 3. -----
>
>       extensions (OCSP, SCT, etc.)
>
> FP: Please consider adding informative references.
>
> [ns] Added
>
> Zaheduzzaman Sarker's No Objection
>
> Thanks for the work on this document. I found it well written and I have
> minor
> comments and Nits
>
> Comment :
>   * As this document asked for a IANA registration entry with DTLS-OK,
> hence
>   this mechanism is OK to be used with DTLS. I understand the heavily
>   references to TLS 1.3 as it relay on the mechanisms described there.
> However,
>   I found it odd not find any reference to DTLS1.3 (we had it on the last
>   formal IESG telechat, it is quite ready to be referenced). Is this
>   intentional? is it supposed to be that this mechanism defined in this
>   document on can be used with DTLS1.2?
>
> [ns] This was a common concern, the text has been updated.
>
>   * Section 7.3 & 7.4: is "active connection" defined somewhere? it would
> be
>   good if some descriptive texts are added for clarification as done for
> the
>   other bullets in the same list.
>
> [ns] Wording adjusted.
>
>   * For the API considerations I was expecting a API to generate the
>   certificate_request_context.
>
> [ns] An API would likely just return a random number, which isn’t very
> useful. Not defining such an API gives the application freedom to
> choose to use this value in some other way. I don’t have strong
> opinions either way.
>
>
> Nits:
>  * Post-handshake authentication is not defined in section 4.6.3 of TLS 1.3
> [ns] Updated in prior change.
>  * Section 4 & 5: likely copy paste error -- s/as its as its/as its
> [ns] Updated in prior change.
>
> Lars Eggert's No Objection
>
> All comments below are very minor change suggestions that you may choose to
> incorporate in some way (or ignore), as you see fit. There is no need to
> let me
> know what you did with these suggestions.
>
> Section 1, paragraph 9, nit:
> -    TLS (or DTLS) version 1.2 [RFC5246] [RFC6347]. or later are REQUIRED
> -                                       -         -
> +    TLS (or DTLS) version 1.2 [RFC5246][RFC6347] or later are REQUIRED
> [ns] Updated
>
>
> Éric Vyncke's No Objection
>
> Thank you for the work put into this document.
>
> Please find below some non-blocking COMMENT points (but replies would be
> appreciated), and some nits.
>
> I hope that this helps to improve the document,
>
> Regards,
>
> -éric
>
> == COMMENTS ==
>
> I am trusting the WG and the responsible AD for ensuring that the specific
> section in references are the correct ones.
>
> -- Section 2 --
> Unsure where the "terminology" part is...
>
> Also, should "Client" and "Server" be defined here ? Reading section 3, I
> am
> still unsure whether the "client" and "server" are the "TLS client" and
> "TLS
> server" ...
>
> In the same vein, a small figure with all parties involved would be
> welcome by
> the reader.
>
> [ns] Terminology updated to reference the terms in TLS 1.3. There are
> only two parties in this protocol, so I’m not sure a diagram will
> help.
>
> == NITS ==
>
> In some places, 'TLS13' is used as an anchor rather than 'RFC8446'.
> [ns] Corrected.
>
> Sometimes, the document uses 'octet' and sometimes 'bytes'.
> [ns] Interesting. This is likely an oversight due to RFC 8446’s use of
> both terms. I’ve adjusted the text to try to be more consistent with
> that document (octets for on-the-wire, bytes for string values).
>
> -- Section 1 --
> Spurious '.' in the last §.
> [ns] Fixed in previous change.
>
> -- Section 4 --
> "as its as its" ?
> [ns] Fixed in previous change.
>
>
> Ben Kaduk's IESG review + Yes
>
> You can view, comment on, or merge this pull request online at:
>   https://github.com/tlswg/tls-exported-authenticator/pull/73
> Commit Summary
> Editorial suggestions from IESG review
> ----------------------------------------------------------------------
> COMMENT:
> ----------------------------------------------------------------------
>
> I've made some (mostl) editorial suggestions in a github PR:
> https://github.com/tlswg/tls-exported-authenticator/pull/73
>
> Despite the overall YES ballot, I do have some substantive comments that
> may be worth some further consideration.
>
> Do we want to give any insight or examples into how an application might
> choose to use the certificate_request_context?
>
> [ns] I put a proposed line here.
>
> Thanks to Yaron Sheffer for the multiple secdir reviews.  My
> understanding is that the intent of this document is to only allow
> "server_name" to appear in the ClientCertificateRequest structure that
> otherwise uses the "CR" notation in the TLS 1.3 column of the TLS
> ExtensionType Values registry, but not to allow for "server_name" in
> authenticator CertificateRequests or handshake CertificateRequests.
> Assuming that's correct, the easiest way to indicate that would be if
> there was a "Comment" column in the registry that could say "only used
> in ClientCertificateRequest certificate requests and no other
> certificate requests", but there is not such a column in the registry.
> I think it could also work to be much more clear in the IANA
> considerations of this document as to why the "CR" is being added and
> what restrictions there are on its use, even if that's not quite as
> clean of a solution.
>
> [ns] I agree that this is an open question.
>
> Section 1
>
>    multiple identities -  Endpoints that are authoritative for multiple
>       identities - but do not have a single certificate that includes
>       all of the identities - can authenticate additional identities
>       over a single connection.
>
> IIUC this is only new functionality for server endpoints, since client
> endpoints can use different certificates for subsequent post-handshake
> authentication events.
>
> [ns] Correct.
>
>    TLS (or DTLS) version 1.2 [RFC5246] [RFC6347]. or later are REQUIRED
>    to implement the mechanisms described in this document.
>
> Please clarify whether this is a minimum TLS version required for using
> EAs, or that this is a binding requirement on implementations of the
> named protocol versions.  (If the latter is intended we may have some
> discussions about an Updates: tag...)
>
> [ns] I meant “minimum version of 1.2” and have changed the text to reflect
> this.
>
> Section 5.1
>
> Using an exporter value as the handshake context means that any client
> authentication from the initial handshake is not cryptographically
> digested into the exporter value.  In light of what we ran into with
> EAP-TLS1.3, do we want to revisit whether we're still happy with that?
> We should in practice still benefit from the implicit confirmation that
> anything the server participates in after receiving the Client Finished
> means the server properly validated it and did not abort the handshake,
> but it's a bit awkward, especially since the RFC 8446 post-handshake
> authentication does digest the entire initial handshake transcript.
> Would we feel any safer if an exporter variant was available that
> digested the entire initial handshake transcript and we could use that?
>
> [ns] Addressed in the open question section above.
>
>
> Section 5.2.1
>
>    Certificates chosen in the Certificate message MUST conform to the
>    requirements of a Certificate message in the negotiated version of
>    TLS.  In particular, the certificate chain MUST be valid for the
>    signature algorithms indicated by the peer in the
>    "signature_algorithms" and "signature_algorithms_cert" extension, as
>    described in Section 4.2.3 of [TLS13] for TLS 1.3 or the
>    "signature_algorithms" extension from Sections 7.4.2 and 7.4.6 of
>    [RFC5246] for TLS 1.2.
>
> This seems awkward.  Do "the requirements of a Certificate message in
> the negotiated version of TLS" include the actual message structure?
> Because the rest of the document suggests that we always use the TLS 1.3
> Certificate, but this line would set us up for using a TLS 1.2
> Certificate.
>
> [ns] They do not include the actual message structure. The constraints
> are about the certificate_list. I’ve updated the text.
>
>
> START
>
>
> Also, per §1.3 of RFC 8446, "signature_algorithms_cert" also applies to
> TLS 1.2, so the last sentence seems incorrect or at least incomplete.
>
> [ns] Updated
>
> Section 5.2.1
>
>    Otherwise, the Certificate message MUST contain only extensions
>    present in the TLS handshake.  Unrecognized extensions in the
>    authenticator request MUST be ignored.
>
> At least the first sentence seems redundant with the previous section
> (but I did not propose removing this in my editorial PR).  The second
> sentence arguably follows from the RFC 8446 requirements, though I don't
> mind mentioning it again as much as I do for the first sentence.
>
> [ns] I don’t mind being explicit here considering there was some
> confusion about this previously.
>
> Section 5.2.2
>
>    The authenticator transcript is the hash of the concatenated
>    Handshake Context, authenticator request (if present), and
>    Certificate message:
>
>    Hash(Handshake Context || authenticator request || Certificate)
>
>    Where Hash is the authenticator hash defined in section 4.1.  If the
>    authenticator request is not present, it is omitted from this
>    construction, i.e., it is zero-length.
>
> This construction is injective, because the handshake messages start
> with a message HandshakeType.  Should we say so explicitly in the
> document?
>
> [ns] I’m not sure I understand what injective means in this context or
> how it helps the reader.
>
>    If the party that generates the exported authenticator does so with a
>    different connection than the party that is validating it, then the
>    Handshake Context will not match, resulting in a CertificateVerify
>    message that does not validate.  This includes situations in which
>    the application data is sent via TLS-terminating proxy.  Given a
>    failed CertificateVerify validation, it may be helpful for the
>    application to confirm that both peers share the same connection
>    using a value derived from the connection secrets before taking a
>    user-visible action.
>
> Is it safe to reuse the Handshake Context itself as the "value derived
> from the connection secrets"?
>
> [ns] Yes, it should be safe. Updated.
>
> Section 7.4
>
>    It returns the identity, such as the certificate chain and its
>    extensions, and a status to indicate whether the authenticator is
>    valid or not after applying the function for validating the
>    certificate chain to the chain contained in the authenticator.
>
> I strongly recommend not returning the identity in the case when
> validation fails.
>
> [ns] Fair. Updated.
>
>    The API should return a failure if the certificate_request_context of
>    the authenticator was used in a previously validated authenticator.
>
> Is the intent for this to be a "distinct previously validated
> authenticator" (i.e., that the API returns the same value when given the
> same inputs subsequently)?  That does not seem to be the meaning
> conveyed by the current text.
>
> [ns] Yes, this is about distinct validators. Updated.
>
> Section 8.2
>
>    IANA is requested to add the following entries to the registry for
>    Exporter Labels (defined in [RFC5705]): "EXPORTER-server
>    authenticator handshake context", "EXPORTER-client authenticator
>    finished key" and "EXPORTER-server authenticator finished key" with
>
> Don't we also need "EXPORTER-client authenticator handshake context"?
>
> [ns] Yes, added.
>
> Section 9
>
> I think we should more explicitly incorporate the TLS 1.3 security
> considerations by reference.
>
> [ns] Yes.
>
> Section 11.1
>
> I don't think [QUIC-TLS] or RFCs 7250 and 7540 are normative references.
>
> [ns] Updated
>
>
> Murray Kucherawy's No Objection
>
> Why are the SHOULDs in Section 4 only SHOULDs?  Why might an implementer
> reasonably not do what this says?
>
> [ns] Addressed after discussion on list and in reply above.
>
> Similarly, the SHOULDs in Section 7 seem awkward, as they have little to do
> with interoperability.
>
> [ns] It’s true that this is about interoperability. I’m open to
> changing these to lower-case shoulds since they reflect implementation
> recommendations rather than interoperability.
>
> - Nick
>
> _______________________________________________
> TLS mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/tls
>