Re: Comments on wrap/unwrap

On Wed, Jun 12, 2013 at 12:41 PM, Ryan Sleevi <sleevi@google.com> wrote:

> On Wed, Jun 12, 2013 at 12:24 PM, Mark Watson <watsonm@netflix.com> wrote:
> >
> >
> > On Wed, Jun 12, 2013 at 12:00 PM, Ryan Sleevi <sleevi@google.com> wrote:
> >>
> >> On Tue, Jun 11, 2013 at 8:31 AM, Mark Watson <watsonm@netflix.com>
> wrote:
> >> > All,
> >> >
> >> > First, if you receive two versions of this email, they should be
> >> > equivalent.
> >> > I wrote one last night and now find no sign of it in my email system,
> so
> >> > this is just a rewrite and retransmission :-(
> >> >
> >> > Second, sorry for the delay in responding on this topic. I was
> >> > unexpectedly
> >> > out of the office all last week.
> >> >
> >> > I have two comments on the wrap/unwrap implementation in the latest
> >> > draft.
> >> >
> >> > 1) The procedures refer to the import/export procedures which are
> >> > currently
> >> > unspecified. For the mechanism to meet the original requirements, it's
> >> > important that import and export respect the attributes of the Key
> >> > (usage,
> >> > exportable). We need to define mappings for these to/from JWK and in
> the
> >> > import case the values in the JWK must override those in the import
> >> > method
> >> > parameters.
> >>
> >> I'm of a mixed opinion as to whether or not this needs to be specified
> >> to this level. The security model assumed here is already
> >> implementation dependent - whether or an implementation respects these
> >> fields, or provides any other guarantees, is already opaque to the web
> >> application.
> >
> >
> > You could also say that about the extractable attribute, or potentially
> > anything in the spec. The spec defines what a compliant implementation
> must
> > do and there is some value in that even when you don't have 100% proof
> that
> > the UA is compliant.
> >
> > To support the use-case, this needs to be specified as proposed.
> >
> >>
> >>
> >> >
> >> > 2) There is one important difference between our proposal and what
> Ryan
> >> > has
> >> > implemented. We proposed methods that would create/consume a
> >> > JWE-encrypted
> >> > JWK object in one step. Ryan's proposed methods are simpler, more
> >> > flexible
> >> > and agnostic as to the wrapping format. Implementing JWE-encrypted JWK
> >> > takes
> >> > two steps but the JS could also implement other formats.
> >> >
> >> > However, in order to obtain the same properties, in terms of key
> >> > extractibility, as the original proposal it is necessary for any
> >> > temporary
> >> > keys (such as JWE Content Encryption Key) to be encoded in a way that
> >> > carries attributes. To explain this, first note that Ryan's methods
> wrap
> >> > and
> >> > unwrap are effectively compound methods based on the existing encrypt,
> >> > decrypt, import and export: wrap = export+encrypt and
> >> > unwrap=decrypt+import
> >> > where the intermediate result (at the '+') is not exposed to the
> script.
> >> >
> >> > To implement JWE-encrypted JWK with these methods we proceed as
> follows:
> >> >
> >> > Wrap a key X using wrapping key Y:
> >> > (W1) Generate a Content Encryption Key with usage "wrap"
> >> > (W2) Use the CEK with the wrap method to export+encrypt X (using JWK
> >> > format)
> >> > (W3) Use Y and with the wrap method to export+encrypt the CEK
> >> > (W3) Construct the JWE object using the output of steps W2 and W3
> >> >
> >> > Unwrap a key X using wrapping key Y:
> >> > (U1) Extract the wrapped CEK and the payload from the JWE structure
> >> > (U2) Use Y and the unwrap method to decrypt+import the wrapped CEK,
> >> > resulting in a Key object K[CEK] for the CEK
> >> > (U3) Use K[CEK] and the unwrap method to decrypt+import the payload,
> >> > resulting in a Key object for X
> >> >
> >> > Let's suppose the X has extractable=false. In order for this property
> to
> >> > be
> >> > maintained it's essential the K[CEK] is not extractible and does not
> >> > have
> >> > usage "decrypt". Otherwise, it would be possible for the Javascript to
> >> > use
> >> > K[CEK] to simply decrypt the payload, exposing the key X.
> >> >
> >> > Therefore, the wrapped encoding of the CEK must carry the extractible
> >> > and
> >> > usages attributes (for cases where the extractibility property of X
> must
> >> > be
> >> > maintained).
> >>
> >> I don't really see this as the case.
> >>
> >> Again, you're working from a threat model where you don't trust the
> >> executing JS. Obviously, this is a threat model that I disagree with,
> >> but you can mitigate this in an implementation-dependent way that
> >> matches your implementation-dependent dependency on named keys.
> >
> >
> > We can't, because this is an issue also without named keys.
> >
> > Regarding the security model, we previously established that this is an
> > issue and that it affects the extractable attribute as much as these
> other
> > things. We should keep the specification consistent - if we are going to
> > have the extractable attribute at all then we need to be able to deliver
> > non-extractable wrapped keys, which has the implications I described
> above.
> >
> > If the group decides that the model is invalid, then the extractable
> > attribute can be removed and my points above become moot.
>
> I disagree with your conclusion that supporting extractable requires,
> by definition supporting non-extractable wrapped keys - if anything,
> this highlights the general issue with key wrapping.
>

No, that wasn't what I was getting at.

We have extractable in the specification. It makes sense to have it there
only if you accept that there is some value in hiding keys from the JS. So,
let's just say that as long as extractable is in the specification there is
some value in hiding keys from the JS. We have a separate issue to decide
on that point, but for the moment, extractable is in the specification.

If it is valuable to hide keys from the JS, it's also valuable to deliver
wrapped keys into the UA in such a way that they remain hidden from the JS.
That was the main point of our wrap/unwrap proposal which the group agreed
should be implemented in the specification, albeit as a "feature at risk".


> Your proposed solution suggests that the JWE/CEK should be modified to
> support attributes - which helps for the JWE/CEK case, but does
> nothing for other formats, which would conceivably share the same
> issue.
>

No, *your* proposed implementation of wrap/unwrap - in which we require two
operations to implement JWE-encapsulated JWK - implies that the CEK must
carry attributes to maintain the properties of the original proposal (to
which the group agreed).

Other formats may also have the same problem: if the wrapping operation is
a single operation, then there is no need for the temporary keys that must
necessarily be used when wrapping using public key cryptography to carry
separate attributes. It is the splitting into two operations, which is
certainly elegant in the sense of decoupling from the specific wrapping
format, that introduces the requirement for attributes on the temporary
key. At least in the model where you don't which to expose the wrapped key
to the JS.

What we end up with is a solution which can support any wrapping format
(good), but only wrapping formats with attributed-carrying temporary keys
will be able to keep keys hidden from the JS.

I'm fine with this, btw.


>
> CryptoAPI/CNG, for example (and +cc Vijay for the possibility of me
> misstating here) lacks the functionality you're requesting, and has
> worked out quite fine for applications.
>
> PKCS#11 somewhat supports the notion you're trying for, by virtue of
> the CKA_WRAP_TEMPLATE and CKA_UNWRAP_TEMPLATE attribute that applies
> to the wrap/unwrapping key. In the wrap case, the CKA_WRAP_TEMPLATE
> specifies a list of attributes that the *key to be wrapped* must
> match. In the unwrap case, the CKA_UNWRAP_TEMPLATE specifies a list of
> attributes that must not conflict with the caller-supplied list of
> attributes.
>
> My statement about the unwrapping key rejecting extractable=false is a
> failure mode consistent with this latter part. The issue you have
> seems to be whether or not the ability to mutate those attributes
> necessarily needs to be exposed to application authors, or whether it
> can be kept behind the notion of 'implementation dependent' - for
> example, coupled to named keys.
>

Yes, this is another option, in which the unwrapping key carries additional
attributes not just about its own extractability/usages but about the
extractability/usages of keys that it unwraps. We would require a new
UnwrappingKey subclass that exposed these attributes, a way to set them
when generating keys and a mapping of these attributes to JWK. We need this
functionality for the case where there are no named origin-specific keys
(the TOFU case).

I don't feel this is especially elegant. It seems unnecessarily complex
when the solution requiring the format to support attributes on the CEK is
fine for the use-cases at hand.

Another possibility, which is even less elegant but also less complex, is
that when unwrapping keys in a format without attributes, the attributes of
the unwrapping key are simply copied to the unwrapped key (I did say this
was even less elegant ...).


>
> I don't think it's absolutely necessary for non-extractable unwrapped
> keys to be supported in V1.


I do. It was almost the whole point of the original proposal. Let's see
what the rest of the group thinks.

...Mark


> I've provided a proposal that would allow
> implementations that view it as necessary to do so in a manner
> consistent with the generic algorithm, and would prefer to leave it
> 'unspecified' / 'not supported' for the generic case. We can then
> revisit different approaches of how to do so - and exactly what the
> security model of such applications are.
>
> >
> >>
> >>
> >> For example, if unwrapKey() is called where unwrappingKey is a key
> >> from some 'named key store', your implementation could choose to fail
> >> the operation if the calling application ever set extractable=true or
> >> set 'decrypt' in keyUsages. That is, your implementation checks these
> >> while importing.
> >
> >
> > The intention of the named origin-specific keys is that there are no
> special
> > implementation requirements other than the existence of the keys
> themselves
> > and compliance to the WebCrypto specifications.
> >
> > Again, though, this is an issue without named origin-specific keys.
> >
> > ...Mark
> >
> >>
> >>
> >> If you trust the JS executing at that particular time, then you just
> >> set extractable=false and don't add a decrypt key usage.
> >>
> >> >
> >> > There are two issues with this:
> >> >
> >> > (A) This is incompatible with the current JWE specification, where the
> >> > wrapped CEK must be encoded as raw octets (though there is a proposal
> to
> >> > allow JWK here)
> >> >
> >> > (B) If the wrapping key is an RSA key, then the size of the encoded
> CEK
> >> > is
> >> > constrained by the size of object this RSA key can encrypt. The only
> >> > attribute-supporting encoding, JWK, has arbitrary size, so it's not
> >> > clear
> >> > how this constraint works in practice
> >> >
> >> > These issues may both be acceptable. I haven't done the maths for (B)
> to
> >> > see
> >> > what the CEK size constraint is for a given RSA key size. If this
> >> > constraint
> >> > is an issue we would need a more compact key+attributes format than
> JWK.
> >> >
> >> > ...Mark
> >
> >
>