Re: [TLS] New Version Notification for draft-friel-tls-over-http-00.txt

["Owen Friel (ofriel)" <[e-mail]>]

From: Ben Schwartz [mailto:[e-mail]]
Sent: Tuesday 31 October 2017 21:17
To: Owen Friel (ofriel) <[e-mail]>
Cc: Richard Barnes <[e-mail]>; <[e-mail]> <[e-mail]>
Subject: Re: [TLS] New Version Notification for draft-friel-tls-over-http-00.txt



On Tue, Oct 31, 2017 at 5:03 PM, Owen Friel (ofriel) <[e-mail]<mailto:[e-mail]>> wrote:


From: TLS [mailto:[e-mail]<mailto:[e-mail]>] On Behalf Of Ben Schwartz
Sent: 31 October 2017 01:35
To: Richard Barnes <[e-mail]<mailto:[e-mail]>>
Cc: <[e-mail]<mailto:[e-mail]>> <[e-mail]<mailto:[e-mail]>>
Subject: Re: [TLS] New Version Notification for draft-friel-tls-over-http-00.txt

On Mon, Oct 30, 2017 at 7:02 PM, Richard Barnes <[e-mail]<mailto:[e-mail]>> wrote:
It requires awareness in the following sense: If by chance the client is in a nice, open network and the base TLS connection goes directly to the server, CONNECT is kind of unnatural; you would want the client to do something different in that case.

Surely this is equally true/untrue of ATLS.  Why do double-TLS if it can be avoided?  But then, how does the application know whether to do ATLS encapsulation?  It's the same question in both cases.

[ofriel] The draft does state “As an optimisation, clients may choose to only use ATLS as a fallback
   mechanism if certificate validation fails on the transport layer TLS
   connection to the service
”
It should be easy for a device to detect the presence of a middlebox if the network layer TLS connection presents a service certificate that has the expected SAN/CN, but is signed by an unexpected/untrusted CA (i.e. one not baked into/explicitly configured on the device).

Yes, but the client could equally well fall back to HTTP CONNECT in this case.  My comments are all directed to the question of why to use ATLS instead of an existing solution, such as "TLS over HTTP CONNECT".

[ofriel] I fully agree that if a TLS tunnel opened via HTTP CONNECT is possible, then ATLS is not needed as the TLS connection is e2e between C->S and the middlebox is not inspecting traffic. However, if the middlebox proxy enforces TLS inspection and does not allow opening a TLS tunnel via HTTP CONNECT, then the dumb IoT device will fail cert validate on the TLS connection terminated on the middlebox.

  You're correct that you *could* configure the server to handle connect properly, but all of the options for doing this are kind of cumbersome -- either you have to stick a possibly-unnecessary proxy in front of the server, or handle CONNECT on the server, which is not really well-supported by web application frameworks.  By contrast, running data over POST is ubiquitous.

This makes a certain amount of sense to me.  If it were up to me, rather than design a TLS-specific transport, I'd be more inclined to propose a standardized version of something like Crowbar<https://github.com/q3k/crowbar>.
[ofriel] /me reads. ‘like’ appears to be the operative word here based on author comments on https://github.com/q3k/crowbar: “Crowbar DOES NOT PROVIDE ANY DATA CONFIDENTIALITY”

Yes, but "TLS over Crowbar" does provide confidentiality, and that is the alternative to ATLS that I am proposing in this sentence.

[ofriel] On browsing what crowbar code actually does, this makes sense. The crowbar solution includes a crowbar-forward agent that client apps talk to, and this sends the packets received over TCP from the client app inside HTTP bodies to a crowbard server, which in turn forwards them to the target server host:port. It also includes some funky URI and message body prefixes in https://github.com/q3k/crowbar/blob/master/proto.go and some rudimentary HMAC(pwd) based authentication.

While crowbar does support listening for any TCP packet stream (which could be a TLS stream) and forwarding them inside HTTP bodies (via crowbar-forward and crowbard server) to a target host:port; we are getting into the realms of requirements here. What we had set out to achieve was e2e encryption between a client and a service across a TLS intercepting middlebox, by simple reuse of the standard TLS stack, without requiring any extra services. More on this below..

(Or just document that reverse proxies and frameworks ought to do something reasonable with CONNECT.
[ofriel] CONNECT would just open a tunnel to get packets through the proxy to the service, but would require the proxy to *not* attempt to do TLS interception, which is exactly what we are trying to allow. If policy dictates that everything must be intercepted, this mechanism enables that.

I don't understand which proxy you're referring to, or how this distinguishes ATLS from "TLS over HTTP CONNECT".

[ofriel] I hope I clarified this above. TLS over HTTP CONNECT is e2e network layer TLS from C->S, the middlebox is not inspecting content, and ATLS isn’t needed if the middlebox allows this.

)  Otherwise this seems to be ossifying the proxy, privileging TLS and preventing deployment of Noise protocol<http://noiseprotocol.org/> or whatever the future may hold.
[ofriel] One of the reason for blindly transporting TLS and not in any way restricting or customising the TLS records transferred was to be future compatible with all future versions of TLS; and also to allow an application to leverage a single software library for both network transport and application crypto exchanges.

OK, but you can be even more abstracted by simply transporting a byte-stream, and letting the application developer choose what protocol to speak in that stream.  This shares even more code between the "network" and "application" usage of TLS, since they both flow over bytestream abstractions (no need to expose the TLS record layer).

[ofriel] Another packaging of content we had considered, and which is equally valid, is to just send the raw TLS Records in the body and specify Content-Type as something like application/atls+octet-stream, not explicitly include the TLS “session” identifier in the message body and have it set by the service in a suitable Set-Cookie.

Note that the TLS record layer isn’t really exposed per-se to the application in either case. The TLS stacks (OpenSSL and JSSE at least) produce raw byte streams that happen to be TLS records, but the application has no idea that they actually are TLS records. In fact, the stacks don’t even expose APIs directly to consumers for parsing the byte streams produced even if the app wanted to grok the TLS records.

One option for making the transport more generic is to simply include raw bytestreams (SSH, TLS, whatever) in HTTP bodies and define suitable application/xxxx+octet-stream.
It was also pointed to me off-list that you can generate POST requests from Javascript in XHR, but not CONNECT requests.  So doing this over POSTs also makes it accessible to web apps.  (`emscripten libssl.a` left as an exercise to the reader.)

It seems your threat model assumes an adversary who is an active intermediary in your HTTP session.  If so, then this wouldn't seem to protect the user against the threat.
[ofriel] Can you clarify why this doesn’t protect against an active intermediary in the HTTP session? Transferring the TLS records payloads in HTTP bodies is directly analogous to  transferring TLS records over untrusted TCP network transport.

In a "web app", the emscripten-compiled copy of libssl would necessarily be transmitted over the insecure HTTP connection that also contains ATLS.  An active intermediary therefore could, for example, replace the library with a modified version that uses a broken PRNG.

[ofriel] Ok. That’s true for webapps downloading JavaScript over the HTTP connection. But not true for devices that would be using the TLS stack baked into their firmware. An active intermediary couldn’t mess with the ATLS stack in that instance.



--Richard


On Mon, Oct 30, 2017 at 6:43 PM, Ben Schwartz <[e-mail]<mailto:[e-mail]>> wrote:
I don't understand why ATLS allows the app to be less "aware" than HTTP CONNECT.  I also don't understand how an ATLS client is closer to "one code path" than HTTP CONNECT.  It seems to me that your description of client behavior applies equally to ATLS and HTTP CONNECT.

On Mon, Oct 30, 2017 at 6:38 PM, Richard Barnes <[e-mail]<mailto:[e-mail]>> wrote:
But I agree, it would be good to have some more clarity around use cases and why not other solutions.

On Mon, Oct 30, 2017 at 6:37 PM, Richard Barnes <[e-mail]<mailto:[e-mail]>> wrote:
HTTP CONNECT is not great for some use cases because it requires the app to be aware that it's dealing with a proxy.  It's simpler if you can just have one code path that works whether your TLS is intermediated or not.  With the solution outlined in the draft, you can just always ignore the certificate the server sends in the first TLS connection (because it might be from a MitM), and then do all your cert validation, pin checks, etc. at the application layer.

On Mon, Oct 30, 2017 at 6:26 PM, Ben Schwartz <[e-mail]<mailto:[e-mail]>> wrote:
Why not use HTTP CONNECT?  Or rather, it would be helpful to have a section on when/why one would do this vs. CONNECT.

On Mon, Oct 30, 2017 at 6:17 PM, Richard Barnes <[e-mail]<mailto:[e-mail]>> wrote:
Hey TLS folks,

Owen, Max, and I have been kicking around some ideas for how to make secure connections in environments where HTTPS is subject to MitM / proxying.

The below draft lays out a way to tunnel TLS over HTTPS, in hopes of creating a channel you could use when you really need things to be private, even from the local MitM.

Feedback obviously very welcome.  Interested in whether folks think this is a useful area in which to develop an RFC, and any thoughts on how to do this better.

Thanks,
--Richard


On Mon, Oct 30, 2017 at 3:47 PM, <[e-mail]<mailto:[e-mail]>> wrote:

A new version of I-D, draft-friel-tls-over-http-00.txt
has been successfully submitted by Owen Friel and posted to the
IETF repository.

Name:           draft-friel-tls-over-http
Revision:       00
Title:          Application-Layer TLS
Document date:  2017-10-30
Group:          Individual Submission
Pages:          20
URL:            https://www.ietf.org/internet-drafts/draft-friel-tls-over-http-00.txt
Status:         https://datatracker.ietf.org/doc/draft-friel-tls-over-http/
Htmlized:       https://tools.ietf.org/html/draft-friel-tls-over-http-00
Htmlized:       https://datatracker.ietf.org/doc/html/draft-friel-tls-over-http-00


Abstract:
   Many clients need to establish secure connections to application
   services but face challenges establishing these connections due to
   the presence of middleboxes that terminate TLS connections from the
   client and restablish new TLS connections to the service.  This
   document defines a mechanism for transporting TLS records in HTTP
   message bodies between clients and services.  This enables clients
   and services to establish secure connections using TLS at the
   application layer, and treat any middleboxes that are intercepting
   traffic at the network layer as untrusted transport.  In short, this
   mechanism moves the TLS handshake up the OSI stack to the application
   layer.




Please note that it may take a couple of minutes from the time of submission
until the htmlized version and diff are available at tools.ietf.org<http://tools.ietf.org>.

The IETF Secretariat


_______________________________________________
TLS mailing list
[e-mail]<mailto:[e-mail]>
https://www.ietf.org/mailman/listinfo/tls