Re: Extractability

On Thu, Aug 1, 2013 at 12:23 AM, Mark Watson <watsonm@netflix.com> wrote:
> I'll reply more generally later, but quickly regarding the relationship of
> our use-case to TLS: we have designed and deployed a secure application
> protocol using WebCrypto which meets our particular application security
> needs when running over plain http. This includes confidentiality and mutual
> authentication. Such protocols are presumably exactly the kind of thing
> WebCrypto is intended to be used for and so not in any sense unreasonable.
> It's also not "technically unrealistic" since we have done it.
>
> Clearly, when scripts are delivered over HTTP the level of trust that one
> might have in the Javascript will be substantially less than one might have
> in the browser. Nevertheless, even when HTTPS is used it's obvious that
> script and browser are different things, subject to different attacks by
> different actors. It's obvious that they "might" be subject to different
> levels of trust. But whether they *are* subject to such different levels of
> trust, and the implications of this, are highly application dependent.
>
> Our requirement to work without TLS arises from the need to access content
> at HTTP URLs as explained in my earlier mail.

While non-normative, we've included the following text in the Security
Considerations since the 2013-01-08 WG Working Draft (
http://www.w3.org/TR/2013/WD-WebCryptoAPI-20130108/#security-developers
)

"While this API provides important functionality for the development
of secure applications, it does not try to address all of the issues
that may arise from the web security model. As such, application
developers must take care to ensure against common attacks such as
script injection by making use of appropriate security functionality
such as Content Security Policy and the use of TLS."

I suspect that where we differ is a fundamental difference on where
the security boundary exists. Speaking as an implementor, our goal has
always been to enable "Web Applications" to take advantage of this.
That is, the security boundary inherently includes the JavaScript
being executed.

If I can correctly distill your example, your goal is to authenticate
the User Agent itself (which seems consistent with your support for
pre-provisioned, origin-specific named keys), without necessarily
concerning yourself about the JavaScript being executed.

Presumably, the UA uses some out-of-bands mechanism to authenticate
the JS, because you simply cannot argue there is confidentiality and
mutual authentication *of* the JS without first establishing the
provenance of the JS, which you cannot use WebCrypto to do.

Is this a correct summary of your implementation?

>
> ...Mark
>
>
> On Tue, Jul 30, 2013 at 8:50 PM, Ryan Sleevi <sleevi@google.com> wrote:
>>
>> Virginie,
>>
>> I apologize, but I'm not sure I entirely follow your response.
>>
>> We're certainly agreed that extractability is a desirable feature, and
>> has been since the beginning. We also agree that supporting key
>> wrap/unwrap are desirable features, and being able to define the
>> extractability of keys that are wrapped/unwrapped are likewise
>> entirely reasonable. I've only seen one member of the WG suggest
>> otherwise, and only when building an unrealistic strawman argument.
>>
>> However, your points on the trust model are a bit confusing for me.
>> The core of the technical discussion here has been precisely upon what
>> is an acceptable trust model - whether or not the existing web trust
>> model that exists for all other Web/JS APIs is sufficient, or whether
>> it's our responsibility in the WG to attempt to define something
>> beyond that.
>>
>> The *key* question for this discussion is whether or not the API can
>> or should presume 'trust' in the validity and veracity of the script
>> being executed. All other Web APIs presume exactly that. WebAppSec/Web
>> Security provide the tools and techniques to establish that further -
>> eg: HTTPS, CORS, CSP, etc. Are you suggesting this is not the case -
>> and that we must face some additional presumptions, or that we cannot
>> rely on the vast body of work that came before this WG?
>>
>> I again position that there is absolutely no value in the API - either
>> the inputs or the outputs - if you don't have some degree of assurance
>> that the script is executing. I have argued, from the very beginning,
>> that attempts to build an alternative to TLS *on the general web*
>> through the use of this API are both unreasonable requirements and
>> technically unrealistic, given the very web security model defined and
>> expanded upon by WebAppSec (and, to a lesser degree, DAP/SysApps).
>> With respect to the SysApps model, I acknowledge it's a very different
>> threat model - one that provides even *more* guarantee about the
>> veracity of the script being executed, ergo the concerns raised by
>> Mark are arguably even less pressing.
>>
>> This is not a position that it somehow specific to the Netflix
>> proposal - I've expressed this position for other cases (eg:
>> http://lists.w3.org/Archives/Public/public-webcrypto/2013Apr/0088.html
>> ), and we've seen a similar position by Arun when editing the use
>> cases and considering the 'threat model' of what was colloquially
>> termed the "Facebook Use case"
>>
>> I'm hoping you can clarify as to what you mean when you say "This is
>> not the job of the web crypto WG do define this trust model" - since
>> this trust model so clearly affects and influences the requirements
>> and design.
>>
>> Cheers,
>> Ryan
>>
>> On Mon, Jul 29, 2013 at 1:35 AM, GALINDO Virginie
>> <Virginie.GALINDO@gemalto.com> wrote:
>> > Hello Mark, Ryan, and all,
>> >
>> >
>> >
>> > I need obviously to get into the details of your mail exchanges, but in
>> > order that we focus on the right tasks in the working group, I would
>> > like to
>> > sumup my understanding of the situation :
>> >
>> > -          The extractability attribute is something the WG would like
>> > to
>> > have,
>> >
>> > -          The trust models between javascript and browser are different
>> > –
>> > even if in the end both of them are breakable, efforts you involve to
>> > break
>> > it are different,
>> >
>> > -          This is not the job of the web crypto WG do define this trust
>> > model, as we mentioned that security model would be a general work
>> > treated
>> > in collaboration with WebAppSec and Web Security IG.
>> >
>> >
>> >
>> > Note that if we fail to manage the extractability in border cases (such
>> > as
>> > wrap/unwrap/import/export) then we can think either about dropping the
>> > extractability or dropping the border case functions. In both cases, our
>> > deliverable will miss the opportunity to be really valuable and answer a
>> > real market demand.
>> >
>> >
>> >
>> > So lets work on a technical solution, not to have this happening.
>> >
>> >
>> >
>> > Regards,
>> >
>> > Virginie
>> >
>> >
>> >
>> >
>> >
>> > From: Mark Watson [mailto:watsonm@netflix.com]
>> > Sent: samedi 27 juillet 2013 10:56
>> > To: Ryan Sleevi
>> > Cc: Harry Halpin; GALINDO Virginie; public-webcrypto@w3.org
>> > Subject: Extractability
>> >
>> >
>> >
>> > All,
>> >
>> >
>> >
>> > I changed the subject of this thread, because Ryan is raising again the
>> > question of whether the extractable attribute makes sense at all. Or,
>> > rather, the more general question of whether use-cases where the
>> > Javascript
>> > is less trusted by the service than the UA is in scope for our work.
>> >
>> >
>> >
>> > I believe we should decide on this more general question and then
>> > consider
>> > the implications for extractable and for wrap/unwrap.
>> >
>> >
>> >
>> > On that question, I don't believe it is our job in WebCrypto to perform
>> > detailed application security analysis. We are providing tools to
>> > application security engineers and the tools we provide are based on
>> > use-cases.
>> >
>> >
>> >
>> > For our use-case, we cannot use HTTPS to deliver our page because we
>> > have to
>> > access non-SSL resources such as CDN content. Switching CDNs to SSL is
>> > expensive in terms of cost, computing resources, and network overhead.
>> > And,
>> > we no longer want to use SSL because our target devices do not and
>> > cannot
>> > get accurate time. (There are a number of CE devices that wish to use
>> > the
>> > HTML5 solution -- the standard has to consider them and not just desktop
>> > browsers.). In fact our whole reason for using WebCrypto is to build our
>> > own
>> > secure application protocol to use instead of SSL.
>> >
>> >
>> >
>> > As a result, MITM attacks against the Javascript are relatively easy.
>> >
>> >
>> >
>> > Further, in the Netflix case then there is value in attacking the crypto
>> > to
>> > extract the keying material because that allows you to bypass Netflix
>> > service restrictions or deny service to the legitimate user. If keying
>> > material is known to be non-extractable (modulo TOFU), then we can still
>> > be
>> > assured that it is the same browser we are talking to (or at least a
>> > browser
>> > to which the same user has migrated the keys, if they are sophisticated
>> > enough to do that.)
>> >
>> >
>> >
>> > ...Mark
>> >
>> >
>> >
>> > On Thu, Jul 25, 2013 at 9:42 PM, Ryan Sleevi <sleevi@google.com> wrote:
>> >
>> > On Thu, Jul 18, 2013 at 7:30 PM, Mark Watson <watsonm@netflix.com>
>> > wrote:
>> >>
>> >>
>> >> On Thu, Jul 18, 2013 at 7:09 PM, Ryan Sleevi <sleevi@google.com> wrote:
>> >>>
>> >>> On Mon, Jul 8, 2013 at 5:12 PM, Mark Watson <watsonm@netflix.com>
>> >>> wrote:
>> >>> > Returning to the subject of the original post, and to start off the
>> >>> > discussion.
>> >>> >
>> >>> > Ryan has mentioned two other possibilities for solving this problem,
>> >>> > so
>> >>> > I'd
>> >>> > like to take a moment to describe my understanding of those.
>> >>> >
>> >>> > (1) Implicit unwrap semantics in pre-provisioned keys
>> >>> >
>> >>> > A pre-provisioned key with usage unwrap could be imbued with
>> >>> > behaviors
>> >>> > that
>> >>> > dictate the extractable and usage attributes of keys that it unwraps
>> >>> > or
>> >>> > even
>> >>> > that imbue the unwrapped keys with other such properties. The former
>> >>> > would
>> >>> > be sufficient for "single step" key wrapping, where the final key to
>> >>> > be
>> >>> > used
>> >>> > for encryption, decryption, signature or signature verification is
>> >>> > wrapped
>> >>> > directly with the pre-provisioned key. The special property of the
>> >>> > pre-provisioned key ensures that the final key has extractable =
>> >>> > false.
>> >>> >
>> >>> > If you want to have two steps, for example the key you are
>> >>> > transferring
>> >>> > is
>> >>> > encrypted using a temporary Content Encryption Key (as in JWE) and
>> >>> > then
>> >>> > this
>> >>> > CEK is wrapped using the pre-provisioned key, then you not only need
>> >>> > the
>> >>> > pre-provisioned key to force extractable = false and usage = unwrap
>> >>> > on
>> >>> > the
>> >>> > CEK, but it must also transfer a special property to the CEK, so
>> >>> > that
>> >>> > when
>> >>> > this in turn is used for unwrapping the resultant key always has
>> >>> > extractable
>> >>> > = false.
>> >>>
>> >>> Correct. The "Named Pre-provisioned keys" is already imbued with
>> >>> special properties by definition, so this is consistent.
>> >>>
>> >>> JWK is not unique in this 'two step' form - consider multi-party
>> >>> RSA-KEM - you have the RSA key, the derived per-party KEK, and the
>> >>> shared, protected key.
>> >>>
>> >>> >
>> >>> > (2) Explicit attributes on wrapping keys
>> >>> >
>> >>> > A key with usage "unwrap" also has properties which dictate the
>> >>> > attributes
>> >>> > of keys that it unwraps. Let's call these properties
>> >>> > "unwrap-extractable"
>> >>> > and "unwrap-usages". Whenever a key, W, is used to perform an unwrap
>> >>> > operation, the unwrapped key, K, gets it's attributes set as
>> >>> > follows:
>> >>> >
>> >>> > K.extractable = W.unwrap-extractable
>> >>> > K.usages = W.unwrap-usages
>> >>> >
>> >>> > Again, this is sufficient for single-step unwrapping. When the
>> >>> > wrapping
>> >>> > key
>> >>> > W is generated, the unwrap-extractable and unwrap-usages properties
>> >>> > are
>> >>> > set
>> >>> > to 'false' and the intended usages of the expected wrapped key,
>> >>> > respectively, When it comes to unwrapping the unwrapped key, K, gets
>> >>> > the
>> >>> > appropriate properties.
>> >>>
>> >>> Correct.
>> >>>
>> >>> This matches PKCS#11's CKA_WRAP_TEMPLATE and CKA_UNWRAP_TEMPLATE
>> >>> properties, for which the smart card and secure element industry have
>> >>> long since embraced as sufficient for a variety of high-security needs
>> >>> (eg: eID cards, as a number of members have pointed out)
>> >>>
>> >>> >
>> >>> > However, if the intended usage of the key K is also for unwrapping
>> >>> > (as
>> >>> > in
>> >>> > the two-step key wrapping described above), we need a way to set
>> >>> > K.unwrap-extractable and K.unwrap-usages.
>> >>> >
>> >>> > Theoretically, we could go down the path of having
>> >>> > unwrap-extractable
>> >>> > and
>> >>> > unwrap-usages each be an array, popping the first value on each
>> >>> > unwrap
>> >>> > operation, i.e.
>> >>> >
>> >>> > K.extractable = W.unwrap-extractable[ 0 ]
>> >>> > K.usages = W.unwrap-usages[ 0 ]
>> >>> > K.unwrap-extractable = W.unwrap-extractable[ 1 : ]
>> >>> > K.unwrap-usages = W.unwrap-usages[ 1 : ]
>> >>> >
>> >>> > (using python-like slice notation)
>> >>> >
>> >>> > It may not be necessary to explicitly expose these attributes on the
>> >>> > Key
>> >>> > object: it may be sufficient to have them settable at key creation
>> >>> > time.
>> >>> >
>> >>> > The other option is to have the extractable and usage attributes
>> >>> > carried
>> >>> > securely with the wrapped key, as I have proposed.
>> >>>
>> >>> Note: This solution ONLY works with JWE-protected-JWK keys - it does
>> >>> not and cannot work with 'raw' or 'pkcs8'/spki. The smart card / HSM /
>> >>> SE industry certainly seems to recognize that mixing/matching as you
>> >>> propose only really works in an implementation-specific manner - see
>> >>> the CKM_SEAL_KEY proposal in the OASIS TC to see how the very nature
>> >>> of 'opaque' key blobs is left up to implementations because of this.
>> >>>
>> >>> You missed the third option though - which is that the (JavaScript)
>> >>> caller specifies the policy.
>> >>
>> >>
>> >> As you explain below, that's not an option that maintains the
>> >> extractability
>> >> functionality. In this mail, I was exploring options which do that.
>> >>
>> >>>
>> >>>
>> >>> If I can sum up the discussion so far, the two objections against this
>> >>> last point (eg: what is currently specified) are:
>> >>> 1) It allows end-users to manipulate variables (eg: in the Javascript
>> >>> console) to circumvent this
>> >>> 2) In the event of an XSS, an attacker can unwrap a key and set
>> >>> extractable to false.
>> >>>   2.1) The first attack requires the attacker has previously observed
>> >>> a wrapped key in transit (eg: MITM) before an XSS, then later XSSes
>> >>> and replays the original key with 'extractable' as true.
>> >>>   2.2) The second attack requires the attacker have XSSed the site,
>> >>> the server send a wrapped key, and the XSS change 'extractable' to
>> >>> true.
>> >>>
>> >>> I see #1 as an explicit non-goal for a general web spec - it's a
>> >>> feature, not a bug.
>> >>
>> >>
>> >> I don't see it as consistent with the existing extractable attribute
>> >> though.
>> >> We should be consistent. Following your approach, we should remove the
>> >> extractable attribute (not that I am proposing this).
>> >>
>> >>>
>> >>> #2.1 can (and should) be mitigated via HTTPS and related.
>> >>> #2.2 can (and should) be mitigated via CSP and related.
>> >>
>> >>
>> >> There are many ways in which the Javascript running on the users
>> >> machine
>> >> may
>> >> not be the Javascript that either the user or the service provider
>> >> expects.
>> >
>> > If you think that this is relevant to the threat model, you absolutely
>> > need to provide an expansion on this.
>> >
>> > If you're suggesting the UA defend against "malware", then that's a
>> > non-starter. If you're talking about extensions or other such, then
>> > either the user was informed and consented, or it's malware. I don't
>> > see how you can arrive in a situation where neither party has
>> > authorized something AND that being a situation that we as a WG must
>> > deal with.
>> >
>> >
>> >> The extractability attribute provides some protection against such
>> >> scripts
>> >> obtaining the raw keying material once it has been installed, provided
>> >> the
>> >> browser itself is not compromised. We're not in a position to do
>> >> security
>> >> engineering for every possible application here, we're providing tools
>> >> and
>> >> extractability is a useful one.
>> >>
>> >> Given the above, it's completely reasonable to want to maintain this
>> >> property with wrapped keys.
>> >
>> > Again, if this is the malware case, it's completely unreasonable to
>> > want to maintain this property.
>> >
>> >
>> >>
>> >>>
>> >>>
>> >>> Finally, the Structured Clonability of Key objects permits other
>> >>> creative uses that have strong parallels to existing software such as
>> >>> middleware, for example, by having a 'trusted' origin perform the
>> >>> unwrapping, and then postMessaging() to the untrusted origin (which,
>> >>> for example, may not be able to support strict CSP policies), while
>> >>> still preserving attributes.
>> >>
>> >>
>> >> Sure, but you are making a bunch of assumptions or imposing a bunch of
>> >> constraints on how applications are designed. What I can say is that
>> >> for
>> >> our
>> >> application, this wouldn't work. Our security analysis suggests that we
>> >> should in all cases attach a different level of trust to the Javascript
>> >> code
>> >> than we do to the browser code. Both can be compromised, of course, but
>> >> the
>> >> ways in which the Javascript can be attacked are more numerous and
>> >> varied.
>> >>
>> >> ...Mark
>> >>
>> >
>> > Naturally, I strongly disagree with this as being a reasonable goal for
>> > the
>> > API.
>> >
>> > It is, in my view, unreasonable to simultaneously suggest you 'trust'
>> > JS to perform crypto but then don't trust the JS performing the
>> > crypto. As we discussed from the very beginning, the mere act of
>> > permitting cryptographic operations is often more than sufficient to
>> > leverage any number of attacks - the formal analysis of PKCS#11 we
>> > discussed in our first F2F was very much a demonstration of this and
>> > why such a goal is unreasonable for any generic API.
>> >
>> > You're right, it absolutely makes a statement that "If you're going to
>> > run code, you trust the code you're going to run" - and using *other*
>> > mechanisms to improve or augment that trust (eg: extensions/sysapps,
>> > as we've also discussed extensively).
>> >
>> > As has also been discussed at length, if you're assuming a MITM that
>> > can modify JS, then it's entirely reasonable to assume that if they're
>> > not attacking the crypto, they're attacking any number of other
>> > aspects - including stripping out the crypto entirely. It's a never
>> > ending game of whack-a-mole that benefits no one, compared to actually
>> > dealing with the trust problem where it belongs - with the JS itself.
>> >
>> >
>
>

Received on Tuesday, 6 August 2013 22:41:35 UTC