Blog from January, 2014

Survival in a World of Zero Trust

Try the following Thought Experiment. Take 100 random users, put them in a room, and ask them to log into your favorite federated web app. If you tell them to first go create an account with some IdP of Last Resort, they will probably groan and quite likely not grok the value of what you are trying to demonstrate. Indeed.

Instead tell them to choose their favorite social IdP on the discovery interface. This will immediately win over a sizable proportion of your audience who will blithely log into your app with almost zero effort (since most social IdPs will happily maintain a user session indefinitely).

However, as Jim Fox demonstrated on the social identity mailing list the other day, not all users feel comfortable performing a federated login at a social IdP. Some users have a healthy distrust for social IdPs, and moreover, that lack of trust is on the rise. So be it.

What can we conclude from this Thought Experiment? Here's my take. Bottom line: Federation is Hard. By no means is the Federated Model a done deal. It may or may not survive, and moreover, I can't predict with any accuracy what will prevail.

That said, I believe in the Federated Model and I want it to work in the long run, so here's what I think we should do in the short term. The appearance of social IdPs on the discovery interfaces of Federation-wide SPs is an inevitability. The sooner we do it, the better off we as users will be. For some (many?), this will simplify the federation experience, and we dearly need all the simplification we can get.

We don't need any more IdPs of Last Resort in the wild, at least not until the trust issues associated with IdPs have been worked out. I'm talking of course about multifactor authentication, assurance, user consent, and privacy, all very hard problems that continue to impede the advance of the Federated Model. In today's atmosphere of Zero Trust, it makes absolutely no sense to keep building and relying on password-based SAML IdPs. That elusive One IdP That Rules Them All simply doesn't exist. We need something better. Something that's simple, safe, and private.

If you're still reading this, you'll want to know what the viable alternatives are. Honestly, I don't know. All I can say is that I'm intrigued by the user centric approaches of the IRMA project and the FIDO Alliance. If similar technologies were to proliferate, it would be a death knell for the centralized IdP model. In its place would rise the Attribute Authority, and I don't mean the SSO-based AAs of today. I mean standalone AAs that dish out attribute assertions that end users control. This is the only approach I can see working in a World of Zero Trust.

Certificate Migration in Metadata

In the last few months, InCommon Operations has encountered two significant interoperability “incidents” (but it is likely there were others we didn’t hear about) that were subsequently traced to improper migration of certificates in metadata. When migrating certificates in metadata, site administrators are advised to carefully follow the detailed processes documented in the wiki. These processes have been shown to maintain interoperability in the face of certificate migration, which tends to be error prone.

Of particular importance is the migration of certificates in SP metadata since a mistake here can affect users across a broad range of IdPs (whereas a mistake in the migration of a certificate in IdP metadata affects only that IdP’s users). When migrating a certificate in SP metadata, the most common mistake is to add a new certificate to metadata before the SP software has been properly configured. This causes IdPs to unknowingly encrypt assertions with public keys that have no corresponding decryption keys configured at the SP. A major outage at the SP will occur as a result.

Another issue we’re starting to see is due to increased usage of Microsoft AD FS at the IdP. As it turns out, AD FS will not consume an entity descriptor in SP metadata that contains two encryption keys. To avoid this situation, the old certificate being migrated out of SP metadata should be marked as a signing certificate only, which avoids any issues with AD FS IdPs. See the Certificate Migration wiki page for details.

Please share this information with your delegated administrators. By the way, if you’re not using delegated administration to manage SP metadata, please consider doing so since this puts certificate migration in the hands of individuals closest to the SP.

Social Identity: A Path Forward

Social Identity was a hot topic at Identity Week in San Francisco last November. Some of the participants questioned the wisdom of Social Identity in the first place: Do we really want to give the keys to the kingdom to Google? Wouldn’t it be in the Federation's best interest to continue to encourage campuses to deploy their own IdPs?

I think we can have our cake and eat it too if we properly limit the scope of a centralized Social Gateway for all InCommon participants.

First, what problem are we trying to solve? The primary motivation for a centralized Social Gateway is to attract Federation-wide SPs to join InCommon. Jim Basney says it best: Given the current penetration of InCommon within US higher education (approximately 20% of US HE institutions operate InCommon IdPs), a “catch-all” IdP is essential to provide a complete federation solution. A Social Gateway would go a long way towards filling this gap.

Second, what problems are we trying to avoid? Most importantly, we want campuses to continue their orderly transition to a federated environment, and so a centralized gateway should not be seen as an alternative to deploying a local IdP.

Not surprisingly, there is also a privacy concern. Like most other Social IdPs in this post-Snowden era, Google is seen by many as a privacy risk. To minimize this risk, we can (and should) limit the attributes that transit the gateway regardless of the attributes Google actually asserts. Moreover, note that a social gateway is inherently privacy-preserving in the sense that it masks (from Google) the end SPs the user visits.

With that background, consider the following proposed centralized Google Gateway for all InCommon participants:

  • lightweight deployment
  • reassert email and person name and that’s all; any other attributes asserted by Google are routinely dropped on the Gateway floor
  • manufacture and assert ePPN at the Gateway
  • no extra attributes, no trust elevation, no invitation service

Such a gateway can not be used in lieu of a campus IdP. If a campus wants to go that route, presumably it can deploy a campus-based gateway on its own.

It is well known that a service provider can request attributes by reference using the SAML AuthnRequest Protocol. Usually the reference is to an <md:AttributeConsumingService> element in metadata, but as discussed in a recent blog post on data minimization, metadata is not required. The identity provider may map the reference to an attribute bundle in whatever way makes sense.

Below we describe a non-use case for the <md:AttributeConsumingService> element in the InCommon Federation. In this case, the service provider doesn't request attributes at all. Instead, the identity provider releases a predefined bundle of attributes to all service providers in a class of service providers called an Entity Category.

There is a class of service providers called the Research and Scholarship Category of service providers (SPs). Associated with the Research and Scholarship (R&S) Category is an attribute bundle B. An identity provider (IdP) supports the R&S Category if it automatically releases attribute bundle B to every SP in the R&S Category.

Every SP in the R&S Category is required to have an <md:AttributeConsumingService> element in its metadata. Now suppose an SP encodes attribute bundle A in metadata such that A is a subset of B. In other words, the SP requests fewer attributes than what the IdP has agreed to release. In this case, an IdP can choose to release A and still support the R&S Category.

Although the software supports this behavior, few (if any) IdPs are configured this way. Instead, IdPs release B to all SPs regardless of the attributes called out in SP metadata. Why? Well, it's easier for the IdP to release B across the board but there's another more important reason that depends on the nature of attribute bundle, so let me list the attributes in the bundle along with some sample values:
displayName:Tom Scavo OR (givenName:Tom AND sn:Scavo)

Note that all of the attributes in this bundle are name-based attributes. If you're going to release one, you may as well release all of them since each attribute valu encodes essentially the same information. This is why IdPs choose to release the entire bundle across the board: there are no privacy benefits in releasing a strict subset, and so it's simply easier to release all attributes to all R&S SPs.

I'm sure one could come up with a hypothetical use case for which there are real privacy benefits in releasing subset A of bundle B but we haven't bumped into such a use case yet. In any event, note the following:

  • AFAIK, no software implementation supports more than one <md:AttributeConsumingService> element in metadata so there isn't much point in calling out the index of such an element in the <samlp:AuthnRequest>.
  • Use of the AttributeConsumingServiceIndex XML attribute as described in the blog post is interesting, but entity attributes give the same effect, and moreover, entity attributes are in widespread use today (at least in higher ed).
  • I doubt any IdP in the InCommon Federation would be inclined to implement a liberal attribute release policy such as "release whatever attributes are called out in the <md:AttributeConsumingService> element in metadata" since this is a potentially serious privacy leak.

This leads to the following prediction: the <md:AttributeConsumingService> element in metadata and the AttributeConsumingServiceIndex XML attribute in the <samlp:AuthnRequest> will turn out to be historical curiosities in the SAML protocol. At this point, the best approach to attribute release appears to be the Entity Category (of which the R&S Category is an example).