Awesome

First-Party Sets

This document proposes a new web platform mechanism to declare a collection of related domains as being in a First-Party Set.

A Work Item of the Privacy Community Group.

Editors:

• Kaustubha Govind, Google
• Harneet Sidhana, Microsoft
• Johann Hofmann, Google

Participate

• https://github.com/privacycg/first-party-sets/issues

Table of Contents

• Introduction
• Goals
• Non-goals
• Use Cases
• Applications
• Proposal
  • Defining a "set" through use case-based "subsets"
  • Abuse mitigation measures
  • Leveraging the Storage Access API
  • Extending the Storage Access API
  • Administrative Controls
• UI Treatment
• Domain Schemes
• Clearing Site Data on Set Transitions
  • Examples
• Alternative designs
  • Signed Assertions and set discovery instead of static lists
  • Self-attestation and technical enforcement
  • Origins instead of registrable domains
• Security and Privacy Considerations
  • Avoid weakening new and existing security boundaries
• Prior Art
• Acknowledgements

Introduction

Browsers have proposed a variety of tracking policies and privacy models (Chromium, Edge, Mozilla, WebKit) which scope access to user identity to some notion of first-party. In defining this scope, we must balance two goals: the scope should be small enough to meet the user's privacy expectations, yet large enough to provide the user's desired functionality on the site they are interacting with.

First-Party Sets (FPS) is a web platform mechanism, proposed within the context of browser efforts to phase out support for third-party cookies, through which site authors of multi-domain sites may declare relationships between domains such that the browser may understand the relationships and handle cookie access accordingly.

The core principle of allowing browsers to treat collections of known related sites differently from otherwise unrelated sites is grounded in ideas that had been previously discussed in the W3C (such as Affiliated Domains), the now defunct IETF DBOUND working group, and previously deployed in some browsers (such as the Disconnect.me entities list).

There are two key components to the proposal:

• The framework governing how relationships between domains may be declared, and
• The method by which the browser may manage cross-domain cookie access based on the declared relationship between domains.

Goals

• Allow for browsers to understand the relationships between domains of multi-domain sites such that they can make decisions on behalf of the user and/or effectively present that information to the user.
• Uphold existing web security principles such as the Same Origin Policy.

Non-goals

• Expansion of capabilities beyond what is possible without recent browser-imposed privacy mitigations such as restrictions on third party cookies or cache partitioning.
• Third-party sign-in between unrelated sites.
• Information exchange between unrelated sites for ad targeting or conversion measurement.
• Other use cases which involve unrelated sites.
• Define specific UI treatment.

(Some of these use cases are covered by other explainers from the Privacy Sandbox.)

Use Cases

On the modern web, sites span multiple domains and many sites are owned & operated by the same organization. Organizations may want to maintain different top-level domains for:

• App domains - a single application may be deployed over multiple domains, where the user may seamlessly navigate between them as a single session.
  • office.com, live.com, microsoft.com (reference)
  • lucidchart.com, lucid.co, lucidspark.com, lucid.app (reference)
• Brand domains
  • uber.com, ubereats.com
• Country-specific domains to enable localization
  • google.co.in, google.co.uk
• Common eTLD
  • For example, gov.uk, and service.gov.uk are on the Public Suffix List and have UK government agencies/services as subdomains which get treated as separate registrable domains by browsers; but share services such as consent management that rely on access to cross-domain cookies.
• Sandbox domains that users never directly interact with, but exist to isolate user-uploaded content for security reasons.
  • google.com, googleusercontent.com
  • github.com, githubusercontent.com
• Service domains that users never directly interact with, but provide services across the same organization’s sites. - github.com, githubassets.com - facebook.com, fbcdn.net

Note: The above have been provided only to serve as real-world illustrative assumed examples of collections of domains that are owned by the same organization; and have not all been validated with the site owners.

This proposal anchors on the use cases described above to develop a framework for the browser to support limited cross-domain cookie access. This will allow browsers to ensure continued operation of existing functionality that would otherwise be broken by blocking cross-domain cookies ("third-party cookies"), and will support the seamless operation of functionality such as:

• Sign-in across owned & operated properties
  • bbc.com and bbc.co.uk
    • Websites may also consider using the FedCM API for single sign-on functionality, if the relevant login flows can be encapsulated with the API's supported use cases.
  • sony.com and playstation.com
• Support for embedded content from across owned & operated properties (e.g. videos/documents/resources restricted to the user signed in on the top-level site)
• Separation of user-uploaded content from other site content for security reasons, while allowing the sandboxed domain access to authentication (and other) cookies. For example, Google sequesters such content on googleusercontent.com, GitHub on githubusercontent.com, CodePen on cdpn.io. Hosting untrusted, compromised content on the same domain where a user is authenticated may result in attackers’ potentially capturing authentication cookies, or login credentials (in case of password managers that scope credentials to domains); and cause harm to users.
  • Alternative solution: Sandboxed domains can also consider using partitioned cookies, if their user flows do not involve the sandboxed domain appearing in top-level contexts.
• Shared services, such as consent management across domains with a common eTLD suffix; such as gov.uk. Repeatedly asking for cookie consent on individual gov.uk sites may be confusing to users, erode trust in the website’s functioning, and cause fatigue; because users think of all subdomains as being part of one gov.uk website.
• Analytics/measurement of user journeys across O&O properties to improve quality of services.

Applications

In support of the various browser privacy models, first-party sets only control when embedded content that would otherwise be considered third-party can access its own state. Examples:

• Sites may annotate individual cookies to be sent across same-party, cross-domain contexts by using the proposed SameParty cookie attribute.
• Top-level key for partitioned cookies a.k.a “chips”. This allows third-party sites (such as embedded SaaS providers) to provide access to the same user session across multiple top-level sites within the same first-party set (reference use-case)
• Issuing WebID directed identifiers by First-Party Set, so the same account can be shared across multiple applications or services provided by the same first-party.
• Applying Privacy Budget across an entire First-Party Set, in order to prevent fingerprinting entropy from being accumulated across domains that are able to communicate in an unconstrained manner due to access to cross-domain, same-party cookies.
• Top and/or second level key for cache partitioning, potentially with site opt-in.

Proposal

At a high level, a First-Party Set is a collection of domains, for which there is a single "set primary" and potentially multiple "set members." Only site authors will be able to submit their own set, and they will be required to declare the relationship between each "set member" to its "set primary." This declaration will be grounded in the use cases described above and defined by "subsets."

Defining a "set" through use case-based "subsets"

Throughout the evolution of this proposal, we considered how to define a single boundary that could determine set inclusion. However, formulating a definition or set of criteria that can both acknowledge the complex multi-domain dependence of websites and preserve a limited privacy boundary proved to be challenging. Instead of using a single definition or set of criteria to apply to a range of use cases, we propose granular criteria and handling to be applied by use case by specifying "subsets."

At time of submission, "set primaries" and "set members" will be declared. Set members could include a range of different domain types, matching up to the different types of use cases (or subsets) like ccTLDs or common eTLDs; domains that users never directly interact with, like service or sandbox domains; and domains where users may benefit from a seamless session, like brand or app domains.

We propose enumerating the range of applicable subsets within a set (beginning with subsets that correlate to the use cases described above), requiring that a member domain must meet the definition of a single subset to be part of the set. For example, consider the following table as an example First-Party Sets schema:

Set primary: exampleA.com

While we think this subset framework has the clear benefit of furthering transparency around why a domain has been added to a set, the primary value to this framework is that the browser could handle each subset differently:

Abuse mitigation measures

We consider using a public submission process (like a GitHub repository) to be a valuable approach because it facilitates our goal to keep all set submissions public and submitters accountable to users, civil society, and the broader web ecosystem. For example, a mechanism to report potentially invalid sets may be provisioned. We expect public accountability to be a significant deterrent for intentionally miscategorized subsets.

The following technical checks also help to mitigate abuse:

• Mutual exclusivity to ensure a domain isn't part of multiple First-Party Sets
• .well-known file check on all domains to ensure authorized submissions
• Check against the Public Suffix List to ensure that sets are composed of valid registrable domains

Additionally, there are other enforcement strategies we could consider to further mitigate abuse. If there is a report regarding a domain specified under the "service" subset, potential reactive enforcement measures could be taken to validate that the domain in question is indeed a "service" subset.

For some subsets, like the "associated" subset, objective enforcement may be much more difficult and complex. In these situations, the browser's handling policy, such as a limit of three domains, should limit the scope of potential abuse. Additionally, we think that site authors will be beholden to the subset definition and avoid intentional miscategorization as their submissions would be entirely public and constitute an assertion

of the relationship between domains.

Leveraging the Storage Access API

To facilitate the browser's ability to handle each subset differently, we are proposing leveraging the Storage Access API (SAA) to enable cookie access within a FPS.

With the SAA, sites may actively request cross-site cookie access, and user-agents may make their own decisions on whether to automatically grant or deny the request or choose to prompt the user. We propose that browsers supporting FPS incorporate set membership information into this decision. In other words, browsers may choose to automatically grant cross-site access when the requesting site is in the same FPS, or in a particular subset of the same FPS, as the top-level site.

We'd like to collaborate with the community in evolving the Storage Access API to improve developer and user experience and help the SAA better support the use cases that FPS is intended to solve. One way to do that is through extending the API surface in a way that makes it easier for developers to use the SAA without integrating iframes:

Extending the Storage Access API

SAA currently requires that the API: (a) be invoked from an iframe embedding the origin requesting cross-site cookies access, and that (b) the iframe obtains user activation before making such a request. We anticipate that the majority of site compatibility issues (specifically, those that FPS intends to address) involve instances where user interaction within an iframe is difficult to retrofit, e.g. because of the usage of images or script tags requiring cookies. Additionally, since cross-site subresources may be loaded synchronously by the top-level site, it may be difficult for the subresources to anticipate when asynchronous cookie access via SAA is granted. To address this difficulty, we propose an extension to the SAA that we hope will make it easier for developers to adopt this change.

Note: Both Firefox and Safari have run into these issues before and have solved them through the application of an internal-only "requestStorageAccessForOrigin" API (4, 5), that is applied on a case-by-case basis by custom browser scripts (Safari: 6, 7 Firefox: 8, 9, 10).

As we continue to flesh out this proposal, we invite feedback from browser vendors, web developers, and members of the web community. We will continue engagement through issues in this repo and through discussions in WICG.

Administrative controls

For enterprise usage, browsers typically offer administrators options to control web platform behavior. UA policy is unlikely to cover private domains, so browsers might expose administrative options for locally-defined first-party sets.

UI Treatment

In order to provide transparency to users regarding the First-Party Set that a web page’s top-level domain belongs to, browsers may choose to present UI with information about the First-Party Set owner and the members list. One potential location in Chrome is the Origin/Page Info Bubble - this provides requisite information to discerning users, while avoiding the use of valuable screen real-estate or presenting confusing permission prompts. However, browsers are free to choose different presentation based on their UI patterns, or adjust as informed by user research.

Note that First-Party Sets also gives browsers the opportunity to group per-site controls (such as those at chrome://settings/content/all) by the “first-party” boundary instead of eTLD+1, which is not always the correct site boundary.

Domain Schemes

In accordance with the Fetch spec, user agents must "normalize" WebSocket schemes to HTTP(S) when determining whether a particular domain is a member of a First-Party Set. I.e. ws:// must be mapped to http://, and wss:// must be mapped to https://, before the lookup is performed.

User agents need not perform this normalization on the domains in their static lists; user agents may reject static lists that include non-HTTPS domains.

Clearing Site Data on Set Transitions

Sites can change which First-Party Set they are a member of, for example through acquisition or divestiture. Since membership in a set could provide access to cross-site cookies via automatic grants of the Storage Access API, we need to pay attention to these transitions so that they don’t link user identities across all the FPSs they’ve historically been in. In particular, we must ensure that a domain cannot transfer a user identifier from one First-Party Set to another when it changes its set membership. While a set member may not always request and be granted access to cross-site cookies, for the sake of simplicity of handling set transitions, we propose to treat such access as always granted.

In order to achieve this, site data needs to be cleared on certain transitions. The clearing should behave like Clear-Site-Data: "*", which includes cookies, storage, cache, as well as execution contexts (documents, workers, etc.). We don’t differentiate between different types of site data because:

• A user identifier could be stored in any of these storage types.
• Clearing just a few of the types would break sites that expect different types of data to be consistent with each other.

Since member sites can only add/remove themselves to/from FPSs with the consent from the owner, we look at first-party set changes as a site changing its FPS owner.

If a site’s owner changed:

1. If this site had no FPS owner, the site's data won't be cleared.
   • Pro: Avoids adoption pain when a site joins a FPS.
   • Con: Unclear how this lines up with user expectations about access to browsing history prior to set formation.
2. Otherwise, clear site data of this site.

Potential modification, which adds implementation complexity:

3. If this site's new owner is a site that previously had the same FPS owner as the first site, the site's data won't be cleared.
   • Pro: Provides graceful transitions for examples (f) and (g).
   • Con: Multi-stage transitions, such as (h) to (i) are unaccounted for.

Examples

a. Site A and Site B create a FPS with Site A as the owner and Site B as the member. Site data will not be cleared.

b. Site C joins the existing FPS as a member site where Site A is the owner. Site data will not be cleared.

c. Given an FPS with owner Site A and members Site B and Site C, if Site D joins this FPS and becomes the new owner; the previous set will be dissolved and the browser will clear data for Site A, Site B and Site C.

d. Given an FPS with owner Site A and members Site B and Site C, if Site B leaves the FPS, the browser will clear site data for Site B.

e. Given two FPSs, FPS1 has owner Site A and members Site B and Site C and FPS2 has owner Site X and member Site Y, if they join together as one FPS with Site A being the owner, the browser will clear site data for Site X and Site Y.

With the potential modification allowing sites to keep their data if the new set owner was a previous member:

f. Given an FPS with owner Site A and members Site B and Site C, if no site is added or removed, just Site C becomes the owner and Site A becomes the member, no site data will be cleared.

g. Given an FPS with owner Site A and members Site B and Site C, if Site A leaves the FPS and Site B becomes the owner, the browser will clear site data for Site A.

h. & i. Given the FPS with owner Site A and member Site B and Site C, if Site D joins this set as a member and later becomes the owner, site data of Site A, Site B and Site C is only preserved if the user happens to visit during the intermediate stage.

Alternative designs

Synchronous cross-site cookie access within same-party contexts

Where a Storage Access API invocation is automatically granted due to membership in the same First-Party Set, a similar effect may be achieved by user agents always allowing cross-site cookie access across sites within the same set. Such cookie access may be subject to rules such as SameSite, and depend on specification of a cookie attribute such as SameParty. This would allow for synchronous cookie access on subresource requests, and, for most part, allows legacy same-party flows to continue functioning with minimal adoption costs involved for site authors. However, it prevents browsers' ability to mediate these flows and potentially intervene on behalf of users. Additionally, Storage Access API is already the preferred mechanism for gaining cross-site cookie access on major browsers such as Safari and Firefox.

Signed Assertions and set discovery instead of static lists

Static lists are easy to reason about and easy for others to inspect. At the same time, they can develop deployment and scalability issues. Changes to the list must be pushed to each user's browser via some update mechanism. This complicates sites' ability to deploy new related domains, particularly in markets where network connectivity limits update frequency. They also scale poorly if the list gets too large.

The Signed Assertions based design proposes an alternative solution that involves the browser learning the composition of sets directly from the websites that the user visits. To prevent privacy risks from personalized sets and ensure policy conformance, they are still verified by an independent entity through a digital signature.

This design is significantly more complex than the consumption of a static list, especially when implementing discovery and fetching of sets in a privacy-preserving manner. As such, we prefer to start with the simpler static list approach, leaving the possibility of introducing a more complex alternative in the future.

Using EV Certificate information for dynamic verification of sets

Extended Validation (EV) Certificates, in addition to backing encrypted exchange of information on the web, require verification of the legal entity associated with the website a certificate is issued for and encode information about this legal entity in the certificate itself. It might be possible to match this information for sites presenting EV certificates (or use the subjectAltName on a single EV certificate) to build First-Party Sets. This could be used in place of Signed Assertions as part of a dynamic set discovery mechanism.

However, such an automatic mechanism would result in a very tight coupling of identity and feature exposure through First-Party Sets to the existing certificate infrastructure.

It's likely that this would negatively impact the deployment and use of encryption on the web, for example by forcing sites to obtain EV certificates as the only way to ensure continued functionality. A revocation of a certificate that is used for FPS would have grave implications (such as deletion of all local data through the Clear Site Data mechanism) and thus complicate the revocation process.

See Issue 12 for an extended discussion.

Self-attestation and technical enforcement

Instead of having a verification entity check conformance to policy; it may be possible to rely on a combination of:

• Self-attestation of UA Policy conformance by submitter.
• Technical consistency checks such as verifying control over domains, and ensuring that no domain appears in more than one set.
• Transparency logs documenting all acceptances and deletions to enable accountability and auditability.
• Mechanism/process for the general public to report potential violations of UA Policy.

However, at this time we do not believe it is possible to enforce against the formation of consortiums of unrelated entities, and thus will require some form of verification entity to guard against that.

Origins instead of registrable domains

A first-party set is a collection of origins, but it is specified by registrable domains, which carries a dependency on the public suffix list. While this is consistent with the various proposed privacy models as well as cookie handling, the security boundary on the web is the origin, not registrable domain.

An alternate design would be to instead specify sets by origins directly. In this model, any https origin would be a possible first-party set owner, and each origin must individually join a set, rather than relying on the root as we do here. For continuity with the existing behavior, we would then define the registrable domain as the default first-party set for each origin. That is, by default, https://foo.example.com, https://bar.example.com, and https://example.com:444 would all be in a set owned by https://example.com. Defining a set explicitly would override this default set.

This would reduce the web's dependency on the public suffix list, which would mitigate various problems. For instance, a university may allow students to register arbitrary subdomains at https://foo.university.example, but did not place university.example on the public suffix list, either due to compatibility concerns or oversight. With an origin-specified first-party set, individual origins could then detach themselves from the default set to avoid security problems with non-origin-based features such as cookies. (Note the __Host- cookie prefix also addresses this issue.)

This origin-defined approach has additional complications to resolve:

• There are a handful of features (cookies, document.domain) which are scoped to registrable domains, not origins. Those features should not transitively join two different sets. For instance, we must account for one set containing https://foo.bar.example.com and https://example.com, but not https://bar.example.com. For cookies, we can say that cookies remember the set which created them and we match both the Domain attribute and the first-party set. Thus if https://foo.bar.example.com sets a Domain=example.com cookie, https://example.com can read it, but not https://bar.example.com. Other features would need similar updates.
• The implicit state should be expressible explicitly, to simplify rollback and deployment, which means first-party set manifests must describe patterns of origins, rather than a simple bounded list of domains. In particular, we should support subtree patterns.
• https://foo.example.com's implicit owner is https://example.com. If https://example.com then forms an explicit set which does not include https://foo.example.com, we need to change https://foo.example.com's implicit state, perhaps to a singleton set.
• This complex set of patterns and implicit behaviors must be reevaluated against existing origins every time a first-party set is updated.
• Certificate wildcards (which themselves depend on the public suffix list) don't match an entire subtree. This conflicts with wanting to express implicit states above.

These complexities are likely solvable while keeping most of this design, should browsers believe this is worthwhile.

Security and Privacy Considerations

Avoid weakening new and existing security boundaries

Changes to the web platform that tighten boundaries for increased privacy often have positive effects on security as well. For example, cache partitioning restricts cache probing attacks and third-party cookie blocking makes it much harder to perform CSRF by default. Where user agents intend to use First-Party Sets to replace or extend existing boundaries based on site or origin on the web, it is important to consider not only the effects on privacy, but also on security.

Sites in a common FPS may have greatly varying security requirements, for example, a set could contain a site storing user credentials and another hosting untrusted user data. Even within the same set, sites still rely on cross-site and cross-origin restrictions to stay in control of data exposure. Within reason, it should not be possible for a compromised site in an FPS to affect the integrity of other sites in the set.

This consideration will always involve a necessary trade-off between gains like performance or interoperability and risks for users and sites. User agents should facilitate additional mechanisms such as a per-origin opt-in or opt-out to manage this trade-off. Site owners should be aware of the potential security implications of creating an FPS and form only the smallest possible set of domains that encompasses user workflows/journeys across an application, especially when some origins in the set opt into features that may leave them open to potential attacks from other origins in the set.

Prior Art

• Firefox's entity list
• draft-sullivan-dbound-problem-statement-02
• Single Trust and Same-Origin Policy v2 and affiliated domains from John Wilander to public-webappsec

Open question(s)

• We are still exploring how CHIPS integrates with First-Party Sets. We are working on technical changes to that design as well, and will share updates when we have a proposal.
• While we've proposed a limit of three domains for the "associated" subset, we seek feedback on whether this would be suitable for ecosystem use cases.
• We may consider expanding the technical checks, where possible, involved in mitigating abuse (e.g., to validate the ccTLD and common eTLD subset categories).

Acknowledgements

• Other members of the W3C Privacy Community Group had previously suggested the use of SAA, or an equivalent API; in place of SameParty cookies. Thanks to @jdcauley (1), @arthuredelstein (2), and @johnwilander (3).
• Browser vendors, web developers, and members of the web community provided valuable feedback during this proposal's incubation in the W3C Privacy Community Group.
• This proposal includes significant contributions from previous co-editor, David Benjamin.
• We are also grateful for contributions from Chris Fredrickson and Shuran Huang.