Re: extensions, continued.. (was: 05/24/2016 WebAuthn Summary

On Tue, Jun 7, 2016, 9:34 AM Rolf Lindemann <rlindemann@noknok.com> wrote:

> I agree, FIDO wants and (IMHO) should be privacy preserving.  And I think
> we all agree on that.
>
> So the ‘what’ (IMHO) is undisputed.
>
>
>
> So we need to discuss the ‘how’.
>
>
>
> > Because clients must be able to drop extensions they don't understand,
> all extensions must be requested by the RP.
>
> I don’t think clients have to.
>
>
>
> What is the threat model here?  A user agent (or client) CANNOT repair a
> broken platform.  In other words: If the underlying platform violates
> user’s privacy (e.g. by directly sending GPS location to some server), the
> user agent cannot protect against that.
>
> The user agent (IMHO) can and should protect against Apps/JavaScripts that
> try to violate user’s privacy.
>
> Similar to a user needing to trust the user agent, the user (to at least
> the same extent) needs to trust the underlying platform – or use a
> different one.
>
Agreed.


>
> The user agent/client should be able to drop requests from the RP for
> privacy violating extensions (according to the user’s preferences).  So if
> the user configures the platform to not expose the location, then I would
> also expect the authenticator to respect that setting.
>

Exactly!

Now, how would the authenticator know about the user's preference? It's not
like the authenticator can turn around and ask the user. Also, we don't
have CTAP messages that would allow the platform to tell the authenticator
about the user preferences. The only mechanism we *do* have is for the
platform (or user agent, or client, or whatever you want to call it) to
help the authenticator out here and not forward the location extension
argument if the user doesn't want the location exposed, and do forward it
if the user is ok with the location exposure. That's the mechanism that we
have to signal to the authenticator the user's intention.

So when you say that you expect the authenticator to respect your wishes,
and we have an authenticator that *wants to* respect those wishes (and lets
agree that we only want to support those kinds of authenticators), what
other mechanism do we have to do that, other than the RP signaling the
extension, the client/platform *understanding* the extension, and then -
depending on the user setting - asking or not asking the authenticator to
including the corresponding data?

Dirk.

>
>
> And this is my recollection of where we ended the discussion in the last
> W3C call, i.e. allowing authenticators to add extensions (in the response)
> and allowing the user agent/client to drop extensions in the request in
> order to reflect the user’s preferences.
>
>
>
> I still trust the Authenticator vendors to only add privacy preserving
> extensions which are added by default (similar to me trusting the platform
> vendors to not violating my privacy in the platform itself).
>
>
>
> Kind regards,
>
>    Rolf
>
>
>
>
>
> *Von:* Dirk Balfanz [mailto:balfanz@google.com]
> *Gesendet:* Montag, 6. Juni 2016 21:46
> *An:* Mandyam, Giridhar; public-webauthn@w3.org
> *Betreff:* Re: extensions, continued.. (was: 05/24/2016 WebAuthn Summary
>
>
>
>
>
> On Sat, Jun 4, 2016 at 6:09 PM Mandyam, Giridhar <mandyam@qti.qualcomm.com>
> wrote:
>
> >Does anyone else remember us discussing (perhaps in the FIDO 2.0 WG)
> these extensions?
>
>
>
> Yes, I do.  But I cannot reproduce the communications that occurred within
> one SDO (in this case the FIDO Alliance) on the mailing list of another
> SDO.  I’m not sure this is relevant to your argument anyways.
>
>
>
> I agree it's not relevant other than these three extensions haven't gotten
> as much scrutiny (at least from this spec editor; perhaps from others, too)
> as the other extensions have.
>
>
>
> >How can the client do its job of protecting user privacy if the
> authenticator is allowed to add data to the assertion that the client
> doesn't understand?
>
> I’ll answer with an example.  Chrome browser supports EME.
> MediaKeySessions involve the exchange of messages between the DRM engine
> (CDM) and license server that the client may not understand, and will
> likely involve unique and privacy-impacting identifiers.
>
>
>
> DRM protocols that expose unique identifiers are not exactly an example of
> clients protecting user privacy. FIDO historically draws the line quite
> differently from where DRM systems draw the line. In fact, FIDO
> historically draws the line even more conservatively than the rest of the
> web does (not even considering DRM). Let me give you an example of that:
> when we first introduced app IDs and facets, it was technically possible
> for two origins, let's say google.com and youtube.com, to collaborate and
> access the same key on an authenticator (something that could come in quite
> handy for those two origins, as you can imagine). Note that absent app IDs
> and FIDO, any two origins on the web, if they choose to collaborate, can
> already track a user and agree on a common identifier (by iframing each
> other, run federation protocols, etc). FIDO at that point decided that we
> didn't want to introduce an *additional* channel for different origins to
> track users; so we changed the app id spec to no longer make it possible
> for exampleA.com and exampleB.com to have access to the same keys from an
> authenticator.
>
>
>
> So, historically speaking, *that's* the bar that FIDO has set (and that,
> presumably, we want to inherit here in the webauthn WG), not the bar set by
> DRM systems. DRM systems, to be sure, set the bar differently for good
> reasons - but we're not here to build DRM, we're here to build
> authentication systems that protect the user's security and privacy.
>
>
>
> Answers to your questions below.
>
>
>
>   You can see https://www.w3.org/Bugs/Public/show_bug.cgi?id=26332 for
> more on this topic.  The solution to this was requiring secure context for
> access to EME (
> https://www.chromium.org/Home/chromium-security/deprecating-powerful-features-on-insecure-origins),
> which is what the WebAuth spec already recommends.
>
>
>
> To me, a greater threat to user privacy than unprompted extensions is that
> secure contexts is only a SHOULD requirement in the spec instead of a MUST
> – see https://w3c.github.io/webauthn/#secure-contexts.  But I imagine
> this has already been discussed to death.
>
>
>
> > I therefore propose that these three extensions be changed to included
> a client argument that signals to the authenticator that the information
> described in the extension should be provided by the authenticator.
>
> Qualcomm cannot support the proposal as written.  I would like to pose
> some clarifying questions however:
>
>
>
> a)      Can the client argument be null?
>
> yes.
>
>
>
> b)      What occurs if the client drops an extension that is associated
> with valid client data?  Will the RP know?  e.g. is an exception thrown?
>
> No exceptions is thrown. What happens, and whether the RP will know,
> depends on the extension. Sometimes, the RP won't know whether or not the
> client supported the extension. For example, in the authenticator selection
> extension, the client MUST choose a (random) authenticator if none of the
> specified authenticators are present. So when the RP asks for authenticator
> A, and gets authenticator B, this might be because no authenticator A was
> present or because the client didn't understand the extension - the RP will
> never know. For most extensions, however, I would assume it will be pretty
> obvious from the authenticator response whether the client supported it or
> not.
>
>
>
> c)      What does this mean for unprompted extensions in the packed
> attestation (https://w3c.github.io/webauthn/#sec-raw-data-packed) without
> a corresponding client argument?  It seems that the client cannot just drop
> an unprompted extension and still expect the signature to pass.
>
> Right. "Dropping an unprompted extension" presumably means that the client
> somehow chops off parts of the authenticator response, which will obviously
> break the signature. That's one of the reasons unprompted extensions are a
> bad idea: we can't ask clients to "drop an unprompted extension". We can
> only ask clients to drop extensions that are requested by the RP. Because
> clients must be able to drop extensions they don't understand, all
> extensions must be requested by the RP.
>
>
>
> d)      Similar question for the assertion.
>
>
>
> Same response as above.
>
>
>
> Dirk.
>
>
>
>
>
> -Giri Mandyam
>
>
>
>
>
> *From:* Dirk Balfanz [mailto:balfanz@google.com]
> *Sent:* Saturday, June 04, 2016 11:14 AM
> *To:* Hodges, Jeff <jeff.hodges@paypal.com>; Vijay Bharadwaj <
> vijaybh@microsoft.com>
>
>
> *Cc:* public-webauthn@w3.org
> *Subject:* Re: extensions, continued.. (was: 05/24/2016 WebAuthn Summary
>
>
>
> Hi there,
>
> Until Adam pointed this out to me in Berlin, I had no idea that these
> other three extensions exited. Now, it's certainly my fault - and no one
> else's - for not reading the attestation document more carefully (which is
> where these three extensions were originally defined), but I honestly don't
> remember these three extensions being discussed the way we discussed the
> authenticator-selection and transaction-authorization extensions. Does
> anyone else remember us discussing (perhaps in the FIDO 2.0 WG) these
> extensions?
>
> In particular, I don't understand why they are defined as "unprompted"
> extensions. This is a privacy problem. How can the client do its job of
> protecting user privacy if the authenticator is allowed to add data to the
> assertion that the client doesn't understand? I get Jeff's point about
> innovation if the RP and authenticator can agree on something even if the
> client doesn't know what that something is, but I believe we should err on
> the side of privacy here.
>
> I would also point out that technically speaking, unprompted extensions
> are not allowed according to the current text, which states that "*an
> extension must specify, at minimum, an extension identifier and an
> extension client argument sent via the {{getAssertion()}} or
> {{makeCredential()}} call*".
>
> I therefore propose that these three extensions be changed to included a
> client argument that signals to the authenticator that the information
> described in the extension should be provided by the authenticator.
>
> Thoughts? Opposing or supporting views?
>
> Dirk.
>
>
>
>
> On Fri, May 27, 2016 at 1:07 PM Hodges, Jeff <jeff.hodges@paypal.com>
> wrote:
>
> On 5/27/16, 12:50 PM, "Vijay Bharadwaj" <vijaybh@microsoft.com> wrote:
>
>
>
> You mean you object to allowing the client a say in which extensions are
> emitted? We’re not talking about removing any existing extensions, just
> about clearly defining the circumstances under which an authenticator might
> emit them.
>
>
>
> Yes, we would object to altering the present design that allows for
> authenticators to implement and emit extensions of their own volition, as
> pesently specified (c.f., AAGUID extension, SupportedExtensions extension,
> User Verification Index (UVI) extension).  We feel it is a
> subtle-but-important aspect of fostering the overall ecosystem.
>
>
>
> This entire thread has become quite frayed... having a concrete extension
> proposal on the table may help it coalesce -- I suggest that Giri write up
> the postulated "opaque data" extension using the framework that's presently
> defined in the spec and then hopefully we can more objectively assess it.
>
>
>
> HTH,
>
>
>
> =JeffH
>
>
>
>
>
>
>
>
>
> *From:* Hodges, Jeff [mailto:jeff.hodges@paypal.com
> <jeff.hodges@paypal.com>]
> *Sent:* Friday, May 27, 2016 12:48 PM
> *To:* Vijay Bharadwaj <vijaybh@microsoft.com>
> *Cc:* public-webauthn@w3.org
> *Subject:* Re: extensions, continued.. (was: 05/24/2016 WebAuthn Summary
>
>
>
> On 5/27/16, 12:37 PM, "Vijay Bharadwaj" <vijaybh@microsoft.com> wrote:
>
> One issue with that is that some of the extensions that are currently
> defined (in fact, 3 out of 5) are emitted unprompted by the authenticator.
> Though if we wanted to make this rule, I would be fine with it and we could
> add it in the spec if others agree.
>
>
>
> Essentially the authenticator would still be allowed to ignore requested
> extensions, just not add new ones on its own.
>
>
>
> We paypal object to obviating existing extensions.
>
>
>
>
>
>  *From:* J.C. Jones [mailto:jjones@mozilla.com <jjones@mozilla.com>]
>
> *Sent:* Friday, May 27, 2016 12:33 PM
>
> That's how you'd enforce it: if the authenticator doesn't obey the
> contract, the signature won't be valid when the RP checks it.
>
> Roughly the contract would be: Authenticators will only emit extensions
> they were prompted to emit.
>
>
>
>
>
>
>
>
>
>