Skip to content

Latest commit

 

History

History
713 lines (563 loc) · 32.2 KB

README-v0.5.0.md

File metadata and controls

713 lines (563 loc) · 32.2 KB

Web Access Control (WAC)

Web Access Control (WAC) specification (as used by the Solid project). It is based on Tim Berners-Lee's May 2009 proposal, as originally captured, and subsequently evolved by the community, at Web Access Control Wiki. This spec is a particular subset of the options and extensions described in the wiki.

For use with LDP (and LDP Next) type systems, such as the Solid project (see also the parent spec).

Current Spec version: v.0.5.0 (see CHANGELOG.md)

Table of Contents

  1. Overview
  2. Access Control List Resources
  1. ACL Inheritance Algorithm
  2. Representation Format
  3. Example WAC Document
  4. Describing Agents
  1. Referring to Resources
  2. Modes of Access
  3. Default (Inherited) Authorizations
  4. Not Supported by Design

Overview

Web Access Control (WAC) is a decentralized cross-domain access control system. The main concepts should be familiar to developers, as they are similar to access control schemes used in many file systems. It's concerned with giving access to agents (users, groups and more) to perform various kinds of operations (read, write, append, etc) on resources. WAC has several key features:

  1. The resources are identified by URLs, and can refer to any web documents or resources.
  2. It is declarative -- access control policies live in regular web documents, which can be exported/backed easily, using the same mechanism as you would for backing up the rest of your data.
  3. Users and groups are also identified by URLs (specifically, by WebIDs)
  4. It is cross-domain -- all of its components, such as resources, agent WebIDs, and even the documents containing the access control policies, can potentially reside on separate domains. In other words, you can give access to a resource on one site to users and groups hosted on another site.

Access Control List Resources

In a system that uses Web Access Control, each web resource has a set of Authorization statements describing:

  1. Who has access to that resource (that is, who the authorized agents are)
  2. What types (or modes) of access they have

These Authorizations are either explicitly set for an individual resource, or (more often) inherited from that resource's parent folder or container. In either case, the Authorization statements are placed into separate WAC documents called Access Control List Resources (or simply ACLs).

Containers and Inherited ACLs

The WAC system assumes that web documents are placed in hierarchical containers or folders. For convenience, users do not have to specify permissions on each individual resource -- they can simply set default permissions on a container using a acl:default predicate and the exact container URI as the object, and have all of the resources in that container inherit those permissions.

Individual Resource ACLs

For fine-grained control, users can specify a set of permissions for each individual resource (which overrides any permissions of its parent container). See the Example WAC Document section to get an idea for what that would look like.

ACL Resource Location Discovery

Given a URL for an individual resource or container, a client can discover the location of its corresponding ACL by performing a HEAD (or a GET) request and parsing the rel="acl" link relation.

Example request to discover the location of the ACL resource for a web document at http://example.org/docs/file1 is given below:

HEAD /docs/file1 HTTP/1.1
Host: example.org

HTTP/1.1 200 OK
Link: <file1.acl>; rel="acl"

The request to discover the location of a container's ACL resource looks similar:

HEAD /docs/ HTTP/1.1
Host: example.org

HTTP/1.1 200 OK
Link: <.acl>; rel="acl"

Note that the acl link relation uses relative path URLs (the absolute URL of the ACL resource in the above example would be /docs/.acl).

Clients MUST NOT assume that the location of an ACL resource can be deterministically derived from a document's URL. For example, given a document with a URL of /docs/file1, clients cannot rely on the assumption that an ACL resource exists at /docs/file1.acl, simply using .acl as a suffix. The actual naming convention for ACL resources can differ for each individual implementation (or even for each individual server). If one server locates the ACL resource by appending the suffix .acl, another server could place the ACL resources into a sub-container (locating it at /docs/.acl/file1.acl for the example above).

ACL Schemas

This section is non-normative. ( Issue: solid/specification#169 )

The following schema excerpts use these namespace prefixes:

PREFIX    acl:  <http://www.w3.org/ns/auth/acl#>
PREFIX     dc:  <http://purl.org/dc/terms/>
PREFIX  vcard:  <http://www.w3.org/2006/vcard/ns#>
PREFIX    xsd:  <http://www.w3.org/2001/XMLSchema#>

Authorization Schema

<#authShape> {
  a [acl:Authorization] ;
  acl:accessTo IRI ;
  acl:mode [acl:Read acl:Write acl:Control]+ ;
  (  acl:agent IRI ;
   | acl:agentGroup @<#groupShape> +
  ) ;
}

Group Schema

<#groupShape> {
  a [vcard:Group] ;
  vcard:hasUID IRI /^urn:uuid:/ ;
  dc:created xsd:dateTime ? ;
  dc:modified xsd:dateTime ? ;
  vcard:hasMember IRI + ;
}

ACL Inheritance Algorithm

The following algorithm is used by servers to determine which ACL resources (and hence which set of Authorization statements) apply to any given resource:

  1. Use the document's own ACL resource if it exists (in which case, stop here).
  2. Otherwise, look for authorizations to inherit from the ACL of the document's container. If those are found, stop here.
  3. Failing that, check the container's parent container to see if that has its own ACL file, and see if there are any permissions to inherit.
  4. Failing that, move up the container hierarchy until you find a container with an existing ACL file, which has some permissions to inherit.
  5. The root container of a user's account MUST have an ACL resource specified. (If all else fails, the search stops there.)

It is considered an anti-pattern for a client to perform those steps, however. A client may discover and load a document's individual resource (for example, for the purposes of editing its permissions). If such a resource does not exist, a client SHOULD NOT search for, or interact with, the inherited ACLs from an upstream container.

ACL Inheritance Algorithm Example

Note: The server in the examples below is using the ACL naming convention of appending .acl to resource URLs, for simplicity. As mentioned in the section on ACL discovery, clients should not use or assume any naming convention.

A request (to read or write) has arrived for a document located at /documents/papers/paper1. The server does as follows:

  1. First, it checks for the existence of an individual corresponding ACL resource. (That is, it checks if paper1.acl exists.) If this individual ACL resource exists, the server uses the Authorization statements in that ACL. No other statements apply.
  2. If no individual ACL exists, the server next checks to see if the /documents/papers/ container (in which the document resides) has its own ACL resource (here, /documents/papers/.acl). If it finds that, the server reads each authorization in the container's ACL, and if any of them contain an acl:default predicate, the server will use them (as if they were specified in paper1.acl). Again, if any such authorizations are found, the process stops there and no other statements apply.
  3. If the document's container has no ACL resource of its own, the search continues upstream, in the parent container. The server would check if /documents/.acl exists, and then /.acl, until it finds some authorizations that contain acl:default.
  4. Since the root container (here, /) MUST have its own ACL resource, the server would use the authorizations there as a last resort.

See the Default (Inherited) Authorizations section below for an example of what a container's ACL resource might look like.

Representation Format

The permissions in an ACL resource are stored in Linked Data format (Turtle by default, but also available in other serializations).

Note: A familiarity with Linked Data and RDF Concepts helps with understanding the terminology used in this spec.

WAC uses the http://www.w3.org/ns/auth/acl ontology for its terms. Through the rest of the spec, the prefix acl: is assumed to mean @prefix acl: <http://www.w3.org/ns/auth/acl#> .

Example WAC Document

Below is an example ACL resource that specifies that Alice (as identified by her WebID https://alice.databox.me/profile/card#me) has full access (Read, Write and Control) to one of her web resources, located at https://alice.databox.me/docs/file1.

# Contents of https://alice.databox.me/docs/file1.acl
@prefix  acl:  <http://www.w3.org/ns/auth/acl#>  .

<#authorization1>
    a             acl:Authorization;
    acl:agent     <https://alice.databox.me/profile/card#me>;  # Alice's WebID
    acl:accessTo  <https://alice.databox.me/docs/file1>;
    acl:mode      acl:Read, 
                  acl:Write, 
                  acl:Control.

Describing Agents

In WAC, we use the term Agent to identify who is allowed access to various resources. In general, it is assumed to mean "someone or something that can be referenced with a WebID", which covers users, groups (as well as companies and organizations), and software agents such as applications or services.

Singular Agent

An authorization may list any number of individual agents (that are being given access) by using the acl:agent predicate, and using their WebID URIs as objects. The example WAC document in a previous section grants access to Alice, as denoted by her WebID URI, https://alice.databox.me/profile/card#me.

Groups of Agents

To give access to a group of agents, use the acl:agentGroup predicate. The object of an agentGroup statement is an instance of vcard:Group. The WebIDs of group members are listed in it, using the vcard:hasMember predicate. If a WebID is listed as member of that group, it is given access.

Example ACL resource, shared-file1.acl, containing a group permission:

# Contents of https://alice.databox.me/docs/shared-file1.acl
@prefix  acl:  <http://www.w3.org/ns/auth/acl#>.

# Individual authorization - Alice has Read/Write/Control access
<#authorization1>
    a             acl:Authorization;
    acl:accessTo  <https://alice.example.com/docs/shared-file1>;
    acl:mode      acl:Read,
                  acl:Write, 
                  acl:Control;
    acl:agent     <https://alice.example.com/profile/card#me>.

# Group authorization, giving Read/Write access to two groups, which are
# specified in the 'work-groups' document.
<#authorization2>
    a               acl:Authorization;
    acl:accessTo    <https://alice.example.com/docs/shared-file1>;
    acl:mode        acl:Read,
                    acl:Write;
    acl:agentGroup  <https://alice.example.com/work-groups#Accounting>;
    acl:agentGroup  <https://alice.example.com/work-groups#Management>.

Corresponding work-groups Group Listing document:

# Contents of https://alice.example.com/work-groups
@prefix    acl:  <http://www.w3.org/ns/auth/acl#>.
@prefix     dc:  <http://purl.org/dc/terms/>.
@prefix  vcard:  <http://www.w3.org/2006/vcard/ns#>.
@prefix    xsd:  <http://www.w3.org/2001/XMLSchema#>.

<#Accounting>
    a                vcard:Group;
    vcard:hasUID     <urn:uuid:8831CBAD-1111-2222-8563-F0F4787E5398:ABGroup>;
    dc:created       "2013-09-11T07:18:19Z"^^xsd:dateTime;
    dc:modified      "2015-08-08T14:45:15Z"^^xsd:dateTime;

    # Accounting group members:
    vcard:hasMember  <https://bob.example.com/profile/card#me>;
    vcard:hasMember  <https://candice.example.com/profile/card#me>.

<#Management>
    a                vcard:Group;
    vcard:hasUID     <urn:uuid:8831CBAD-3333-4444-8563-F0F4787E5398:ABGroup>;

    # Management group members:
    vcard:hasMember  <https://deb.example.com/profile/card#me>.

[test against schema]

Group Listings - Implementation Notes

When implementing support for acl:agentGroup and Group Listings, keep in mind the following issues:

  1. Group Listings are regular documents (potentially each with its own .acl).
  2. What authentication mechanism should the ACL checking engine use, when making requests for Group Listing documents on other servers?
  3. Infinite request loops during ACL resolution become possible, if an .acl points to a group listing on a different server.
  4. Therefore, for the moment, we suggest that all Group files which are used for group ACLs are public.

Possible future methods for a server to find out whether a given agent is a member of a group are a matter for future research and possible addition here.

Public Access (All Agents)

To specify that you're giving a particular mode of access to everyone (for example, that your WebID Profile is public-readable), you can use acl:agentClass foaf:Agent to denote that you're giving access to the class of all agents (the general public). For example:

@prefix   acl:  <http://www.w3.org/ns/auth/acl#>.
@prefix  foaf:  <http://xmlns.com/foaf/0.1/>.

<#authorization2>
    a               acl:Authorization;
    acl:agentClass  foaf:Agent;                               # everyone
    acl:mode        acl:Read;                                 # has Read-only access
    acl:accessTo    <https://alice.databox.me/profile/card>.  # to the public profile

Authenticated Agents (Anyone logged on)

Authenticated access is a bit like public access but it is not anonymous. Access is only given to people who have logged on and provided a specific ID. This allows the server to track the people who have used the resource.

To specify that you're giving a particular mode of access to anyone logged on (for example, that your collaborative page is open to anyone but you want to know who they are), you can use acl:agentClass acl:AuthenticatedAgent to denote that you're giving access to the class of all authenticated agents. For example:

@prefix   acl:  <http://www.w3.org/ns/auth/acl#>.
@prefix  foaf:  <http://xmlns.com/foaf/0.1/>.

<#authorization2>
    a               acl:Authorization;
    acl:agentClass  acl:AuthenticatedAgent;                   # everyone
    acl:mode        acl:Read;                                 # has Read-only access
    acl:accessTo    <https://alice.databox.me/profile/card>.  # to the public profile

Note that this is a special case of acl:agentClass usage, since it doesn't point to a Class Listing document that's meant to be de-referenced.

An application of this feature is to throw a resource open to all logged on users for a specific amount of time, accumulate the list of those who access it within that time as a group, and then later restrict access to that group, to prevent spam.

Referring to Origins, i.e. Web Apps

When a compliant server receives a request from a web application running in a browser, the browser will send an extra warning HTTP header, the Origin header.

Origin: https://scripts.example.com:8080

(For background, see also Backgrounder on Same Origin Policy and CORS) Note that the origin comprises the protocol and the DNS and port but none of the path, and no trailing slash. All scripts running on the same origin are assumed to be run by the same social entity, and so trusted to the same extent.

When an Origin header is present then BOTH the authenticated agent AND the origin MUST be allowed access.

As both the user and the web app get to read or write (etc) the data, then they must BOTH be trusted. This is the algorithm the server must go through.

  • If the requested mode is available to the public, then succeed 200 OK with added CORS headers ACAO and ACAH. **
  • If the user is not logged on, then fail 401 Unauthenticated.
  • If the authenticated user is not allowed access, AND the class AuthenticatedAgent is not allowed access, then fail 403 User Unauthorized.
  • If the Origin header is not present, then succeed 200 OK.
  • If the Origin is allowed by the ACL, then succeed 200 OK with added CORS headers ACAO and ACAH.
  • (In future proposed) Look up the owner's webid(s) to check for trusted apps declared there, and if match, succeed 200 OK with added CORS headers ACAO and ACAH.
  • Fail 403 Origin Unauthorized.

Note it is a really good idea to make it clear both in the text of the status message and in the body of the message the difference between the user not being allowed and the web app they are using not being trusted.

** Possible future alternative: Set ACAO header to "*" indicating that the document is public. This will though block in the browser any access made using credentials.

Adding trusted web apps.

** NB: this feature was only added recently and is still considered experimental. It's likely to change in the near future. **

The authorization of trusted web app is a running battle between readers and writers on the web, and malevolent parties trying to break in to get unauthorized access. The history or Cross-Site Scripting attacks and the introduction of the Same Origin Policy is not detailed here, The CORS specification in general prevents any web app from accessing any data from or associated with a different origin. The web server can get around CORS. It is a pain to to do so, as it involves the server code echoing back the Origin header in the ACAO header, and also it must be done only when the web app in question actually is trustworthy.

In solid a maxim is, you have complete control of he data. Therefore it is up to the owner of the data, the publisher, the controller of the ACL, or more broadly the person running the solid server, to specify who gets access, be it people or apps. However another maxim is that you can chose which app you use. So of Alice publishes data, and Bob want to use his favorite app, then how does that happen?

  • A Web server MAY be configured such that a given list of origins is unconditionally trusted for incoming HTTP requests. The origin check is then bypassed for these domains, but all other access control mechanisms remain active.
  • A specific ACL can be made to allow a given app to access a given file or folder of files, using acl:origin.
  • Someone with acl:Control access to the resource could give in their profile a statement that they will allow users to use a given app.
 <#me> acl:trustedApp [ acl:origin  <https://calendar.example.com>;
                        acl:mode    acl:Read, 
                                    acl:Append].
 <#me> acl:trustedApp [ acl:origin  <https://contacts.example.com>;
                        acl:mode    acl:Read, 
                                    acl:Write, 
                                    acl:Control].

We define the owners of the resource as people given explicit Control access to it. (Possible future change: also anyone with Control access, even through a group, as the group can be used as a role)

For each owner x, the server looks up the (extended?) profile, and looks in it for a triple of the form

?x  acl:trustedApp  ?y.

The set of trust objects is the accumulated set of ?y found in this way.

For the app ?z to have access, for every mode of access ?m required there must be some trust object ?y such that

?y  acl:origin  ?z;
    acl:mode    ?m.

Note access to different modes may be given in the same or different trust objects.

Referring to Resources

The acl:accessTo predicate specifies which resources you're giving access to, using their exact URLs as the objects.

Referring to the ACL Resource Itself

Since an ACL resource is a plain Web document in itself, what controls who has access to it? While an ACL resource could in theory have its own corresponding ACL document (for example, file1.acl controls access to file1, and file1.acl.acl could potentially control access to file1.acl), one quickly realizes that this recursion has to end somewhere.

Instead, the acl:Control access mode is used (see below), to specify who has access to alter (or even view) the ACL resource.

Modes of Access

The acl:mode predicate denotes a class of operations that the agents can perform on a resource.

acl:Read

gives access to a class of operations that can be described as "Read Access". In a typical REST API (such as the one used by Solid), this includes access to HTTP verbs GET, and HEAD. This also includes any kind of QUERY or SEARCH verbs, if supported.

acl:Write

gives access to a class of operations that can modify the resource. In a REST API context, this would include PUT, POST, DELETE and PATCH. This also includes the ability to perform SPARQL queries that perform updates, if those are supported.

acl:Append

gives a more limited ability to write to a resource -- Append-Only. This generally includes the HTTP verb POST, although some implementations may also extend this mode to cover non-overwriting PUTs, as well as the INSERT-only portion of SPARQL-based PATCHes. A typical example of Append mode usage would be a user's Inbox -- other agents can write (append) notifications to the inbox, but cannot alter or read existing ones.

acl:Control

is a special-case access mode that gives an agent the ability to view and modify the ACL of a resource. Note that it doesn't automatically imply that the agent has acl:Read or acl:Write access to the resource itself, just to its corresponding ACL document. For example, a resource owner may disable their own Write access (to prevent accidental over-writing of a resource by an app), but be able to change their access levels at a later point (since they retain acl:Control access).

Default (Inherited) Authorizations

As previously mentioned, not every document needs its own individual ACL resource and its own authorizations. Instead, one can can create an Authorization for a container (in the container's own ACL resource), and then use the acl:default predicate to denote that any resource within that container will inherit that authorization. To put it another way, if an Authorization contains acl:default, it will be applied by default to any resource in that container.

You can override the default inherited authorization for any resource by creating an individual ACL just for that resource.

An example ACL for a container would look something like:

# Contents of https://alice.databox.me/docs/.acl
@prefix  acl:  <http://www.w3.org/ns/auth/acl#>.

<#authorization1>
    a                  acl:Authorization;

    # These statements specify access rules for the /docs/ container itself:
    acl:agent          <https://alice.databox.me/profile/card#me>;
    acl:accessTo       <https://alice.databox.me/docs/>;
    acl:mode           acl:Read, 
                       acl:Write, 
                       acl:Control;

    # default says: this authorization (the statements above) 
    #   will also be inherited by any resource within that container 
    #   that doesn't have its own ACL.
    acl:default  <https://alice.databox.me/docs/>.

See also

Background on CORS

Old discussion of access to group files

Group Listings - Authentication of External Requests

This section is not normative

Group Listings via acl:agentGroup links introduce the possibility of an ACL checking engine having to make requests to other servers. Given that access to those external group listings can be protected, the question immediately arises: By what mechanism should the ACL checking engine authenticate its request to external servers?

For example: Alice sends a GET request to a resource on the server https://a.com. The ACL for that resource links to a group listing on an external server, https://b.com. In the process of resolving the ACL, a.com must send a request to b.com, to get that group listing. Note that it's not Alice herself (or her application) that is sending that request, it's actually a.com sending it (as part of trying to resolve its own ACL). How should a.com authenticate itself? Does it have its own credentials, or does it have a way to say that it's acting on behalf of Alice? Or both?

There are several implementation possibilities:

No authentication. The ACL checking engine sends un-authenticated requests to external servers (to fetch group listings). This is the simplest method to implement, but suffers from the limitation that those external group listings need to be public-readable. THIS IS THE ONLY METHOD CURRENTLY IN USE

WebID-TLS Delegation. If your implementation uses the WebID-TLS authentication method, it also needs to implement the ability to delegate its requests on behalf of the original user. (No, the original requester may not be allowed access -- you don't have to able to access a group to be in it) For a discussion of such a capability, see the Extending the WebID Protocol with Access Delegation paper. One thing to keep in mind is - if there are several hops (an ACL request chain across more than one other domain), how does this change the delegation confirmation algorithm? If the original server is explicitly authorized for delegation by the user, what about the subsequent ones?

ID Tokens/Bearer Tokens. If you're using a token-based authentication system such as OpenID Connect or OAuth2 Bearer Tokens, it will also need to implement the ability to delegate its ACL requests on behalf of the original user. See PoP/RFC7800 and Authorization Cross Domain Code specs for relevant examples.

Infinite Request Loops in Group Listings

Since Group Listings (which are linked to from ACL resources using the acl:agentGroup predicate) reside in regular documents, those documents will have their very own .acl resources that restrict which users (or groups) are allowed to access or change them. This fact, that .acls point to Group Listings, which can have .acls of their own, which can potentially also point to Group Listings, and so on, introduces the potential for infinite loops during ACL resolution.

Take the following situation with two different servers:

https://a.com                     https://b.com
-------------        GET          ---------------
group-listA        <------        group-listB.acl
    |                                  ^     contains:
    |                                  |     agentGroup <a.com/group-ListA>   
    v                GET               |
group-listA.acl    ------>        group-listB
  contains:
  agentGroup <b.com/group-listB>

The access to group-listA is controlled by group-listA.acl. So far so good. But if group-listA.acl contains any acl:agentGroup references to another group listing (say, points to group-listB), one runs into potential danger. In order to retrieve that other group listing, the ACL-checking engine on https://b.com will need to check the rules in group-listB.acl. And if group-listB.acl (by accident or malice) points back to group-listA a request will be made to access group-listA on the original server https://a.com, which would start an infinite cycle.

To guard against these loops, implementers have several options:

A) Do not allow cross-domain Group Listing resolutions. The simplest to implement (but also the most limited) option is to disallow cross-domain Group Listings resolution requests. That is, the ACL-checking code could detect agentGroup links pointing to external servers during ACL resolution time, and treat those uniformly (as errors, or as automatic "access denied").

B) Treat Group Listings as special cases. This assumes that the server has the ability to parse or query the contents of a Group Listing document before resolving ACL checks -- a design decision that some implementations may find unworkable. If the ACL checking engine can inspect the contents of a document and know that it's a Group Listing, it can put in various safeguards against loops. For example, it could validate ACLs when they are created, and disallow external Group Listing links, similar to option A above. Note that even if you wanted to ensure that no .acls are allowed for Group Listings, and that all such documents would be public-readable, you would still have to be able to tell Group Listings apart from other documents, which would imply special-case treatment.

C) Create and pass along a tracking/state parameter. For each ACL check request beyond the original server, it would be possible to create a nonce-type tracking parameter and pass it along with each subsequent request. Servers would then be able to use this parameter to detect loops on each particular request chain. However, this is a spec-level solution (instead of an individual implementation level), since all implementations have to play along for this to work. See issue solid/solid#8 for further discussion).

D) Ignore this issue and rely on timeouts. It's worth noting that if an infinite group ACL loop was created by mistake, this will soon become apparent since requests for that resource will time out. If the loop was created by malicious actors, this is comparable to a very small, low volume DDOS attack, which experienced server operators know how to guard against. In either case, the consequences are not disastrous.

Other ideas about specifying trusted apps

  • A reader can ask to use a given app, by publishing the fact that she trusts a given app.
<#me> acl:trustsForUse [ acl:origin  <https://calendar.example.com>;
                         acl:mode    acl:Read, 
                                     acl:Append].
<#me> acl:trustsForUse [ acl:origin  <https://contacts.example.com>; 
                         acl:mode    acl:Read, 
                                     acl:Write, 
                                     acl:Control].

A writer could have also more sophisticated requirements, such as that any app Alice wants to use must be signed by developer from a given list, and so on.

Therefore, by pulling the profiles of the reader and/or the writer, and/or the Origin app itself, the system can be adjusted to allow new apps to be added without bad things happening

Not Supported by Design

This section describes some features or acl-related terms that are not included in this spec by design.

Resource Owners

WAC has no formal notion of a resource owner, and no explicit Owner access mode. For convenience/UI purposes, it may be assumed that Owners are agents that have Read, Write and Control permissions.

acl:accessToClass

The predicate acl:accessToClass is not supported.

Regular Expressions

The use of regular expressions in statements such as acl:agentClass is not supported.