1. Introduction
This section is not normative.
As the web platform is extended to enable more useful and powerful applications, it becomes increasingly important to ensure that the features which enable those applications are enabled only in contexts which meet a minimum security level. This document describes threat models for feature abuse on the web (see §4.1 Threat Models) and outlines normative requirements which should be incorporated into documents specifying new features (see §7 Implementation Considerations).
The most obvious of the requirements discussed here is that application code with access to sensitive or private data be delivered confidentially over authenticated channels that guarantee data integrity. Delivering code securely cannot ensure that an application will always meet a user’s security and privacy requirements, but it is a necessary precondition.
Less obviously, application code delivered over an authenticated and confidential channel isn’t enough in and of itself to limit the use of powerful features by non-secure contexts. As §4.2 Ancestral Risk explains, cooperative frames can be abused to bypass otherwise solid restrictions on a feature. The algorithms defined below ensure that these bypasses are difficult and user-visible.
The following examples summarize the normative text which follows:
1.1. Top-level Documents
Top-level documents are secure as long as they don’t have a non-secure opener browsing context. This is a bit convoluted, so let’s go straight to the examples:
http://example.com/
opened in a top-level browsing
context is not a secure context, as it was not delivered over
an authenticated and encrypted channel.
https://example.com/
opened in a top-level browsing
context is a secure context, as it was delivered over
an authenticated and encrypted channel.
If a secure context opens https://example.com/
in a new
window, that new window will be a secure context, as it is both secure on
its own merits, and was opened from a secure context:
If a non-secure context opens https://example.com/
in a new
window, then things are more complicated. The new window’s status depends on
how it was opened. If the non-secure context can obtain a reference to the
secure context, or vice-versa, then the new window is not a secure context.
This means that the following will both produce non-secure contexts:
<a href="https://example.com/" target="_blank">Link!</a> <script> var w = window.open("https://example.com/"); </script>
The link can be broken via the noopener
link relation, meaning
that the following will both produce secure contexts:
<a href="https://example.com/" rel="noopener" target="_blank">Link!</a> <script> var w = window.open("https://example.com/", "", "noopener"); </script>
W3C’s HTML has only an extremely partial port of
the noopener
concept. <https://github.com/w3c/html/issues/523>
1.2. Framed Documents
Framed documents can be secure contexts if they are delivered from potentially trustworthy origins, and if they’re embedded in a secure context. That is:
https://example.com/
opened in a top-level browsing
context opens https://sub.example.com/
in a frame, then
both are secure contexts, as both were delivered over
authenticated and encrypted channels.
If https://example.com/
was somehow able to frame http://non-secure.example.com/
(perhaps the user has
overridden mixed content checking?), the top-level frame would remain
secure, but the framed content is not a secure context.
If, on the other hand, https://example.com/
is framed
inside of http://non-secure.example.com/
, then it is not a secure context, as its ancestor is not delivered over an
authenticated and encrypted channel.
1.3. Web Workers
Dedicated Workers are similar in nature to framed documents. They’re secure contexts when they’re delivered from potentially trustworthy origins, only if their owner is itself a secure context:
If https://example.com/
in a top-level browsing
context runs https://example.com/worker.js
, then
both the document and the worker are secure contexts.
If http://non-secure.example.com/
in a top-level browsing
context frames https://example.com/
, which runs https://example.com/worker.js
, then neither the framed document
nor the worker are secure contexts.
1.4. Shared Workers
Multiple contexts may attach to a Shared Worker. If a secure context creates a Shared Worker, then it is a secure context, and may only be attached to by other secure contexts. If a non-secure context creates a Shared Worker, then it is not a secure context, and may only be attached to by other non-secure contexts.
If https://example.com/
in a top-level browsing
context runs https://example.com/worker.js
as a Shared
Worker, then both the document and the worker are considered secure
contexts.
https://example.com/
in a different top-level
browsing context (e.g. in a new window) is a secure context, so it may
access the secure shared worker:
https://example.com/
nested in http://non-secure.example.com/
may not connect to the secure
worker, as it is not a secure context.
Likewise, if https://example.com/
nested in http://non-secure.example.com/
runs https://example.com/worker.js
as a Shared
Worker, then both the document and the worker are considered non-secure.
1.5. Service Workers
Service Workers are always secure contexts. Only secure contexts may register them, and they may only have clients which are secure contexts.
If https://example.com/
in a top-level browsing
context registers https://example.com/service.js
,
then both the document and the Service Worker are considered secure
contexts.
2. Framework
An environment settings object is considered a secure
context if the algorithm in §3.1 Is settings object a secure context? returns "Secure
", and a non-secure context otherwise.
Likewise, a global object is considered a secure context if its relevant settings object is a secure context.
2.1. Intergration with WebIDL
This section is non-normative.
A new [SecureContext]
attribute is available for operators, which
ensures that they will only be exposed into secure contexts. The
following example should help:
interface ExampleFeature { // This call will succeed in all contexts. Promise <double> calculateNotSoSecretResult(); // This operation will not be exposed to a non-secure context. [SecureContext] Promise<double> calculateSecretResult(); // The same applies here: the operation will not be exposed to a non-secure context. [SecureContext] boolean getSecretBoolean(); }; [SecureContext] interface SecureFeature { // This interface will not be exposed to non-secure contexts. Promise<any> doAmazingThing(); };
Specification authors are encouraged to use this attribute when defining new features.
2.2. Modifications to HTML
2.2.1. Shared Workers
This section is non-normative.
The SharedWorker()
constructor will throw a SecurityError
exception if
a secure context attempts to attach to an Worker which is not a secure context, and if a non-secure context attempts to attach to a
Worker which is a secure context.
The constructor is modified as follows (though the SharedWorker specification remains the normative reference):
-
As the first substep of the
SharedWorker()
constructor’s current step 6.7 ("If worker global scope is notnull
, then run these steps:"), run the following step:-
If the result of executing §3.1 Is settings object a secure context? on the current settings object does not match the result of executing the same algorithm on worker global scope’s relevant settings object, then throw a
SecurityError
exception, and abort these steps.
-
Note: This change landed in WHATWG’s HTML in whatwg/html#1560.
It’s not clear to me how the W3C’s [WEBWORKERS] document is updated. It looks like it’s pulling content from the WHATWG upstream, which means that the PR linked above should flow into it? But that document hasn’t been updated since 2015, so...
2.2.2. Feature Detection
To determine whether a context is capable of making use of features which require secure contexts, a simple boolean attribute is added to the global object:
partial interface WindowOrWorkerGlobalScope { readonly attribute boolean isSecureContext; };
WindowOrWorkerGlobalScope
does not appear to be defined in W3C’s
HTML. For the purposes of that specification, the IDL above could be
interpreted as:
interface GlobalSecureContext { readonly attribute boolean isSecureContext; }; Window implements GlobalSecureContext; WorkerGlobalScope implements WindowBase64;
Filed as w3c/html#522.
The isSecureContext
attribute’s getter
returns true
if §3.1 Is settings object a secure context? returns "Secure
" when executed upon this
global object’s relevant settings object, and false
otherwise.
3. Algorithms
3.1. Is settings object a secure context?
Given an environment settings object (settings) this algorithm
returns "Secure
" if the object represents a context which the user agent
obtained via a secure channel, and "Not Secure
" otherwise.
-
Let global be settings’s global object.
-
If global is a
WorkerGlobalScope
, then:-
For each
Document
(document) in global’s list of the worker’sDocuments
:-
If §3.1 Is settings object a secure context? returns "
Not Secure
" when executed upon document’s relevant settings object, return "Not Secure
".
-
-
-
Assert: global is a
Window
. -
If settings object’s responsible document has an creator browsing context whose creator context security is "
Not Secure
", return "Not Secure
".Note: Since we take account of creator browsing contexts' status, a popups' status depends on how it is opened, as discussed in §1.1 Top-level Documents.
The 'creator context security' concept landed in WHATWG’s HTML in whatwg/html#1561, but doesn’t yet exist in W3C’s HTML. <https://github.com/w3c/html/issues/524>
-
If settings object’s HTTPS state is "
modern
", return "Secure
".Most of the time, this check will be enough to determine whether a particular context was securely delivered. Documents delivered over TLS will have their HTTPS state set, andsrcdoc
Document
s inherit their ancestor’s HTTPS state (as do other kinds of requests which inherit their requestor’s origin: see the basic fetch algorithm for details on some of these [FETCH]).We only continue past this check in order to allow resources delivered from "trustworthy" but unauthenticated locations like
http://127.0.0.1/
to be treated as creating a secure context. -
Let origin be settings object’s origin.
-
If origin is an opaque origin, set origin to the origin of settings object’s creation URL.
Note: We use the origin of the URL here because sandboxed content that is treated as being in a unique origin (e.g.
<iframe sandbox src="http://127.0.0.1/">
) would otherwise be treated as non-trustworthy by §3.2 Is origin potentially trustworthy?. Since sandboxing is a strict reduction in the content’s capabilities, and therefore in the risk it poses, we look at the origin of its URL to determine whether we would have considered it trustworthy had it not been sandboxed. -
If the result of executing the §3.2 Is origin potentially trustworthy? algorithm on origin is not "
Potentially Trustworthy
", then return "Not Secure
". -
Return "
Secure
".
3.2. Is origin potentially trustworthy?
A potentially trustworthy origin is one which a user agent can generally trust as delivering data securely.
This algorithms considers certain hosts, scheme, and origins as potentially
trustworthy, even though they might not be authenticated and encrypted in the
traditional sense. In particular, the user agent SHOULD treat file
URLs and
as potentially trustworthy. In principle the user agent could treat local
files as untrustworthy, but, given the information that is available to
the user agent at runtime, the resources appear to have been transported
securely from disk to the user agent. Additionally, treating such resources as
potentially trustworthy is convenient for developers building an application
before deploying it to the public.
This developer-friendlyness is not without risk, however. User agents which
prioritize security over such niceties MAY choose to more strictly assign
trust in a way which excludes file
.
On the other hand, the user agent MAY choose to extend this trust to other,
vendor-specific URL schemes like app:
or chrome-extension:
which it can
determine a priori to be trusted (see §7.1 Packaged Applications for detail).
Given an origin (origin), the following algorithm returns
"Potentially Trustworthy
" or "Not Trustworthy
" as appropriate.
-
If origin is an opaque origin, return "
Not Trustworthy
". -
Assert: origin is a tuple origin.
-
If origin’s scheme is either "
https
" or "wss
", return "Potentially Trustworthy
".Note: This is meant to be analog to the a priori authenticated URL concept in [MIX].
Note: The origin of
blob:
andfilesystem:
URLs is the origin of the context in which they were created. Therefore, blobs created in a trustworthy origin will themselves be potentially trustworthy. The origin ofdata:
andjavascript:
URLs, on the other hand, is an opaque origin, which will not be considered potentially trustworthy. -
If origin’s host component matches one of the CIDR notations
127.0.0.0/8
or::1/128
[RFC4632], return "Potentially Trustworthy
". -
If origin’s scheme component is
file
, return "Potentially Trustworthy
". -
If origin’s scheme component is one which the user agent considers to be authenticated, return "
Potentially Trustworthy
".Note: See §7.1 Packaged Applications for detail here.
-
If origin has been configured as a trustworthy origin, return "
Potentially Trustworthy
".Note: See §7.2 Development Environments for detail here.
-
Return "
Not Trusted
".
Note: Neither origin’s domain nor port has any effect on whether or not it is considered to be a secure context.
4. Threat models and risks
This section is non-normative.
4.1. Threat Models
Granting permissions to unauthenticated origins is, in the presence of a network attacker, equivalent to granting the permissions to any origin. The state of the Internet is such that we must indeed assume that a network attacker is present. Generally, network attackers fall into 2 classes: passive and active.
4.1.1. Passive Network Attacker
A "Passive Network Attacker" is a party who is able to observe traffic flows but who lacks the ability or chooses not to modify traffic at the layers which this specification is concerned with.
Surveillance of networks in this manner "subverts the intent of communicating parties without the agreement of these parties" and one "cannot defend against the most nefarious actors while allowing monitoring by other actors no matter how benevolent some might consider them to be." [RFC7258] Therefore, the algorithms defined in this document require mechanisms that provide for the privacy of data at the application layer, not simply integrity.
4.1.2. Active Network Attacker
An "Active Network Attacker" has all the capabilities of a "Passive Network Attacker" and is additionally able to modify, block or replay any data transiting the network. These capabilities are available to potential adversaries at many levels of capability, from compromised devices offering or simply participating in public wireless networks, to Internet Service Providers indirectly introducing security and privacy vulnerabilities while manipulating traffic for financial gain ([VERIZON] and [COMCAST] are recent examples), to parties with direct intent to compromise security or privacy who are able to target individual users, organizations or even entire populations.
4.2. Ancestral Risk
The §3.1 Is settings object a secure context? algorithm walks through all the ancestors of a
particular context in order to determine whether or not the context itself is
secure. Why wouldn’t we consider a securely-delivered document in an iframe
to be secure, in and of itself?
The short answer is that this model would enable abuse. Chrome’s
implementation of [WEBCRYPTOAPI] was an early experiment in locking APIs to
secure contexts, and it does not walk through a context’s ancestors. The
assumption was that locking the API to a resouce which was itself delivered
securely would be enough to ensure secure usage. The result, however, was
that entities like Netflix built iframe
- and postMessage()
-based shims
that exposed the API to non-secure contexts. The restriction was little more
than a speed-bump, slowing down non-secure access to the API, but completely
ineffective in preventing such access.
While the algorithms in this document do not perfectly isolate non-secure contexts from secure contexts (as discussed in §5.1 Incomplete Isolation), the ancestor checks provide a fairly robust protection for the guarantees of authentication, confidentiality, and integrity that such contexts ought to ptovide.
4.3. Risks associated with non-secure contexts
Certain web platform features that have a distinct impact on a user’s security or privacy should be available for use only in secure contexts in order to defend against the threats above. Features available in non-secure contexts risk exposing these capabilities to network attackers:
- The ability to read and modify sensitive data (personally-identifying information, credentials, payment instruments, and so on). [CREDENTIAL-MANAGEMENT-1] is an example of an API that handles sensitive data.
- The ability to read and modify input from sensors on a user’s device (camera, microphone, and GPS being particularly noteworthy, but certainly including less obviously dangerous sensors like the accelerometer). [GEOLOCATION-API] and [MEDIACAPTURE-STREAMS] are historical examples of features that use sensor input.
- The ability to access information about other devices to which a user has access. [DISCOVERY-API] and [WEB-BLUETOOTH] are good examples.
- The ability to track users using temporary or persistent identifiers,
including identifiers which reset themselves after some period of time
(e.g.
window.sessionStorage
), identifiers the user can manually reset (e.g. [ENCRYPTED-MEDIA], Cookies [RFC6265], and [IndexedDB]), as well as identifying hardware features the user can’t easily reset. - The ability to introduce some state for an origin which persists across browsing sessions. [SERVICE-WORKERS] is a great example.
- The ability to manipulate a user agent’s native UI in some way which removes, obscures, or manipulates details relevant to a user’s understanding of their context. [FULLSCREEN] is a good example.
- The ability to introduce some functionality for which user permission will be required.
This list is non-exhaustive, but should give you a feel for the types of risks we should consider when writing or implementing specifications.
Note: While restricting a feature itself to secure contexts is critical, we ought not forget that facilities that carry such information (such as new network access mechanisms, or other generic functions with access to network data) are equally sensitive.
5. Security Considerations
5.1. Incomplete Isolation
The secure context definition in this document does not completely
isolate a "secure" view on an origin from a "non-secure" view on the same
origin. Exfiltration will still be possible via increasingly esoteric
mechanisms such as the contents of localStorage
/sessionStorage
, storage
events, BroadcastChannel
, and others.
5.2. localhost
Section 6.3 of [RFC6761] lays out the resolution of localhost.
and
names falling within .localhost.
as special, and suggests that local
resolvers SHOULD/MAY treat them specially. For better or worse, resolvers
often ignore these suggestions, and will send localhost
to the network
for resolution in a number of circumstances. Given that uncertainty, this
document errs on the conservative side by special-casing 127.0.0.1
, but
not localhost
.
6. Privacy Considerations
The secure context definition in this document does not in itself have any privacy impact. It does, however, enable other features which do have interesting privacy implications to lock themselves into contexts which ensures that specific guarantees can be made regarding integrity, authenticity, and confidentiality.
From a privacy perspective, specification authors are encouraged to consider requiring secure contexts for the features they define.
7. Implementation Considerations
7.1. Packaged Applications
A user agent that support packaged applications MAY whitelist specific URL
schemes whose contents are authenticated by the user agent. For example,
FirefoxOS application resources are referred to by a URL whose scheme component is app:
. Likewise, Chrome’s extensions
and apps live on chrome-extension:
schemes. These could reasonably be
considered trusted origins.
7.2. Development Environments
In order to support developers who run staging servers on non-loopback hosts,
the user agent MAY allow users to configure specific sets of origins as
trustworthy, even though §3.2 Is origin potentially trustworthy? would normally return
"Not Trusted
".
7.3. Restricting New Features
This section is non-normative.
When writing a specification for new features, we recommend that authors and editors guard sensitive APIs with checks against secure contexts. For example, something like the following might be a good approach:
-
If the current settings object is not a secure
context, then:
- [insert something appropriate here: perhaps a Promise could be
rejected with a
SecurityError
, an error callback could be called, a permission request denied, etc.].
- [insert something appropriate here: perhaps a Promise could be
rejected with a
Authors could alternatively ensure that sensitive APIs are only exposed to secure contexts by guarding them with the [SecureContext]
attribute:
[SecureContext] interface SensitiveFeature { Promise<double> getTheSecretDouble(); }; // Or: interface AnotherSensitiveFeature { [SecureContext] void doThatPowerfulThing(); };
7.4. Restricting Legacy Features
This section is non-normative.
The list above clearly includes some existing functionality that is currently available to the web over non-secure channels. We recommend that such legacy functionality be modified to begin requiring a secure context as quickly as is reasonably possible [W3C-PROCESS].
-
If such a feature is not widely implemented, we recommend that the specification be immediately modified to include a restriction to secure contexts.
-
If such a feature is widely implemented, but not yet in wide use, we recommend that it be quickly restricted to secure contexts by adding a check as described in §7.3 Restricting New Features to existing implementations, and modifying the specification accordingly.
-
If such a feature is in wide use, we recommend that the existing functionality be deprecated; the specification should be modified to note that it does not conform to the restrictions outlined in this document, and a plan should be developed to both offer a conformant version of the feature and to migrate existing users into that new version.
7.4.1. Example: Geolocation
The [GEOLOCATION-API] is a good concrete example of such a feature; it is widely implemented and used on a large number of non-secure sites. A reasonable path forward might look like this:
-
Modify the specification to include checks against secure context before executing the algorithms for
getCurrentPosition()
andwatchPosition()
.If the current settings object is not a secure context, then the algorithm should be aborted, and the
errorCallback
invoked with acode
ofPERMISSION_DENIED
. -
The user agent should announce clear intentions to disable the API for non-secure contexts on a specific date, and warn developers accordingly (via console messages, for example).
-
Leading up to the flag day, the user agent should announce a deprecation schedule to ensure both that site authors recognize the need to modify their code before it simply stops working altogether, and to protect users in the meantime. Such a plan might include any or all of:
-
Disallowing persistent permission grants to non-secure origins
-
Coarsening the accuracy of the API for non-secure origins (perhaps consistently returning city-level data rather than high-accuracy data)
-
UI modifications to inform users and site authors of the risk
-
-
8. Acknowledgements
This document is largely based on the Chrome Security team’s work on [POWERFUL-NEW-FEATURES]. Chris Palmer, Ryan Sleevi, and David Dorwin have been particularly engaged. Anne van Kesteren, Jonathan Watt, Boris Zbarsky, and Henri Sivonen have also provided very helpful feedback.