Re: WebCrypto Security Analysis

Sent from my iPhone

On Mar 21, 2014, at 11:58 AM, Ryan Sleevi <sleevi@google.com> wrote:




On Fri, Mar 21, 2014 at 11:53 AM, Richard Barnes <rlb@ipv.sx> wrote:

> On Fri, Mar 21, 2014 at 2:38 PM, Ryan Sleevi <sleevi@google.com> wrote:
>>
>> On Fri, Mar 21, 2014 at 11:34 AM, Mark Watson <watsonm@netflix.com>wrote:
>>>
>>>  On Fri, Mar 21, 2014 at 9:51 AM, Ryan Sleevi <sleevi@google.com> wrote:
>>>
>>>>
>>>> On Mar 21, 2014 9:18 AM, "Mark Watson" <watsonm@netflix.com> wrote:
>>>> >
>>>> >
>>>> >
>>>> >
>>>> > On Thu, Mar 20, 2014 at 1:01 PM, Ryan Sleevi <sleevi@google.com>
>>>> wrote:
>>>> >>
>>>> >>
>>>> >>
>>>> >>
>>>> >> On Wed, Mar 19, 2014 at 7:40 AM, Kelsey Cairns <
>>>> kelsey.cairns@inria.fr> wrote:
>>>> >>>
>>>> >>> Dear W3C Crypto API WG,
>>>> >>>
>>>> >>> Here at INRIA we're starting a security analysis on the current
>>>> draft
>>>> >>> of the Crypto API, co-funded by INRIA and W3C. The idea is to try to
>>>> >>> get some results in before the end of the last call period.
>>>> >>
>>>> >>
>>>> >> Could you define what your actual goal is with this security
>>>> analysis?
>>>> >>
>>>> >> Typically, one does a security analysis of a protocol - does it live
>>>> to the expected goals, and provide the expected assurances. WebCrypto
>>>> itself provides many algorithmic building blocks, and (with the exception,
>>>> arguably, of Wrap/Unwrap), doesn't really implement a protocol itself (as
>>>> opposed something like JOSE JWS or XML DSig, which are arguably both
>>>> formats *and* protocols)
>>>> >>
>>>> >>>
>>>> >>>
>>>> >>> Doing analysis of an API spec is a slightly unusual activity, it can
>>>> >>> often lead to conclusions like "if the API is implemented this
>>>> way.."
>>>> >>> or "if the application program uses the API like this.." which can
>>>> >>> seem a bit superficial, but we will aim to produce something
>>>> concrete
>>>> >>> output in terms of implementation advice, test cases for
>>>> >>> implementations, etc.
>>>> >>>
>>>> >>>
>>>> >>> As an example of the kind of things we find, one of the things we
>>>> were
>>>> >>> looking at in the spec this morning was padding oracles on key
>>>> unwrap
>>>> >>> operations. These are common in implementations of PKCS#11, for
>>>> >>> example.. Following the current WebCrypto spec, if you were to
>>>> unwrap a key using
>>>> >>> AES-CBC or RSA PKCS1v1.5, incorrect padding would lead to
>>>> "DataError"
>>>> >>> or " OperationError" respectively. Meanwhile, the error if the
>>>> >>> ciphertext is correctly padded but the key is too long or too short,
>>>> >>> the error is "SyntaxError". The fact that these are different
>>>> *could*
>>>> >>> be enough to allow a network attacker to obtain the encrypted key by
>>>> >>> chosen ciphertext attack, which would be relevant say for use case
>>>> 2.2
>>>> >>> (Protected Document Exchange).
>>>> >>
>>>> >>
>>>> >> Correct. This is a point of extreme tension within the working group
>>>> - whether or not Key Wrapping / Unwrapping can provide security guarantees
>>>> against the host code executing. This was the key of the debate as to
>>>> whether or not to provide these primitives to begin with - or whether a web
>>>> application can polyfill them.
>>>> >>
>>>> >> Individually, I remainly highly suspicious about this. As a
>>>> security-minded individual, I can tell you there are dozens of ways to
>>>> botch this, beyond just algorithm choice. As an editor, I can simply say
>>>> "Please show more about how this is completely broken", so that the WG can
>>>> take a closer look about the security guarantees it's attempting to make,
>>>> and properly evaluate whether or not these APIs belong. I suspect that some
>>>> members will insist they do, unfortunately, so guidance is welcome.
>>>> >>
>>>> >>>
>>>> >>>
>>>> >>> As a first step we were planning to look in more detail at the key
>>>> >>> management subset of the API, but if there are any areas that are of
>>>> >>> specific concern where you'd like us to take a closer look and you
>>>> >>> haven't had time please let us know. All feedback welcome.
>>>> >>>
>>>> >>> Best,
>>>> >>>
>>>> >>> Graham Steel & Kelsey Cairns
>>>> >>
>>>> >>
>>>> >> I think a clear point of use/misuse to examine would be be the
>>>> issues previously discussed in ISSUE-21 (
>>>> https://www.w3.org/2012/webcrypto/track/issues/21 ) . The WG had, in
>>>> the past, discussed requiring SSL/TLS for this API, as well as requiring
>>>> more active mitigations for scripting issues via CSP (
>>>> http://lists.w3.org/Archives/Public/public-webcrypto/2012Aug/0230.html). There were and are some strong objections to this.
>>>> >>
>>>> >> Since part of your sponsorship includes "implementation advice", and
>>>> conclusions like "if the application program uses the API like this", it
>>>> would be interesting to see if INRIA can come up with any proofs of
>>>> security where the code is delivered over unauthenticated connections (eg:
>>>> HTTP)
>>>> >>
>>>> >> My continued assertion is that this is impossible - messages cannot
>>>> be authenticated as coming from a user/UA, rather than a MITM. Likewise,
>>>> under HTTP, a UA/user cannot authenticate messages as coming from the
>>>> server, rather than a MITM. Encryption/Decryption results cannot be
>>>> protected from being shared with Mallory, and that there can be no
>>>> authenticated key exchange without an OOB means. Especially because Mallory
>>>> can modify the JS operating environment, any proofs of correctness of a
>>>> protocol go out the window, because the operating environment for those
>>>> proofs is malleable. In a PKCS#11 world, this would be similar to a
>>>> "hostile token" that has no pre-provisioned aspects.
>>>> >
>>>> >
>>>> > Ryan is right, of course, that security assertions that can be made
>>>> if the content is delivered over https cannot be made if he content is
>>>> delivered over http. However, this does not mean there are no useful
>>>> security assertions for the case where content is delivered over http. It
>>>> would be good to have the nature of the assertions which can be made
>>>> properly investigated and documented.
>>>> >
>>>> > Specifically, most of the assertions that can be made for the http
>>>> case are in the "Trust on First Use" category: if an authentication key is
>>>> agreed between client and server at time X, then the client can be sure at
>>>> time Y that they are talking to the same entity they were talking to at
>>>> time X (which may be A MITM, or may be the intended server, you don't
>>>> know). Likewise the server can be sure they are talking to the same entity
>>>> at time Y as they were at time X (which, again, may be either a MITM or may
>>>> be the client). If you have other reasons to believe there was no MITM at
>>>> time X, such assertions can be useful.
>>>> >
>>>>
>>>> No, they really aren't.
>>>>
>>>> Regardless of how the key got there (and there are plenty of ways to
>>>> screw that up), the fundamental analysis has to look at whether any of the
>>>> assertions can be trusted if they are being processed by untrusted code:
>>>>
>>>> Sign: Did this message come from Alice or Mallory-injecting-script?
>>>>
>>>> Verify: Did this message come from Bob, or Mallory with
>>>> Mallory-injected script saying Bob
>>>>
>>>> Encrypt: Is the Ciphertext sent to Bob the ciphertext that Alice
>>>> intended, or modified by Mallory?
>>>>
>>>> Decrypt: Is the Plaintext processed by Alice what Bob sent in his
>>>> Ciphertext, or is this Mallory?
>>>>
>>>> Wrap: Is this key the actually Alice's key, or is it a key of Mallory?
>>>>
>>>> Unwrap: Is the unwrapped key actually what Bob intended, or is it
>>>> Mallorie's injected?
>>>>
>>>> Mallory can also force an unprovisioned state at any time, so you need
>>>> a way to authenticate that. WebCrypto cannot provide that, such you must
>>>> rely on side-channels - such as Named Key Discovery or TLS.
>>>>
>>> Yep, as I said, the nature of the assertions which *can* be made with
>>> http content delivery are of the kind "This message came from the same
>>> entity (Alice|Mallory) as I agreed keys with at some previous time X."
>>>
>>> Are you disputing this assertion itself, or whether it is useful ?
>>>
>>> ...Mark
>>>
>>
>> I'm disputing the assertion.
>>
>> "I agreed keys with (Alice|Mallory) at some previous time X" - that
>> statement is self-containable
>> "This message came from (Alice|Mallory)"
>>
>> You can't assert at Time Y that the message came from the same party at
>> Time X, if the *code* used to create that message is delivered insecurely.
>>
>> The point being you may have agreed upon keys with Alice, but then the
>> message came from Mallory - because Mallory injected her code to create a
>> custom message using Alice's credentials.
>>
>> Likewise, when you agree upon keys (at Time X), you can't be sure whether
>> you're agreeing with Alice or Mallory, unless you're using a secure
>> transport.
>>
>> So without a secure transport at both Time X and Time Y, you can't be
>> sure that the party you agreed upon keys with is the same party that
>> authored the message using those keys. Which is the point.
>>
>
> One nuance that might be worth noting though:
>
> If you've marked the key with extractable == false, you at least know that
> you're talking to the same *device* at time Y as at time X.  (Modulo things
> like key extraction/cloning below the JS layer, which aren't part of our
> threat model.)
>
> I'm not sure how useful that property is given that there many be
> Mallory's code running on that device, but...
>
> --Richard
>
>
No, you don't have that guarantee - as the spec is clear to call out that
UA's are free to store the key however they want.

If we're talking only in the context of "Mallory, the remote attacker",
then sure, you have a guarantee of the same UA,


Which, as it happens, is one of the useful guarantees we are interested in
for our service, even without anything else.

...Mark

but not necessarily the same origin (or same application code), since as
you note, Mallory may have postMessage'd the key to herself for later use,
and can always send requests to the "Victim" at will later.