Subresource Integrity

1. Introduction

Sites and applications on the web are rarely composed of resources from only a single origin. For example, authors pull scripts and styles from a wide variety of services and content delivery networks, and must trust that the delivered representation is, in fact, what they expected to load. If an attacker can trick a user into downloading content from a hostile server (via DNS [RFC1035] poisoning, or other such means), the author has no recourse. Likewise, an attacker who can replace the file on the Content Delivery Network (CDN) server has the ability to inject arbitrary content.

Delivering resources over a secure channel mitigates some of this risk: with TLS [TLS], HSTS [RFC6797], and pinned public keys [RFC7469], a user agent can be fairly certain that it is indeed speaking with the server it believes it’s talking to. These mechanisms, however, authenticate only the server, not the content. An attacker (or administrator) with access to the server can manipulate content with impunity. Ideally, authors would not only be able to pin the keys of a server, but also pin the content, ensuring that an exact representation of a resource, and only that representation, loads and executes.

This document specifies such a validation scheme, extending two HTML elements with an integrity attribute that contains a cryptographic hash of the representation of the resource the author expects to load. For instance, an author may wish to load some framework from a shared server rather than hosting it on their own origin. Specifying that the expected SHA-384 hash of https://example.com/example-framework.js is Li9vy3DqF8tnTXuiaAJuML3ky+er10rcgNR/VqsVpcw+ThHmYcwiB1pbOxEbzJr7 means that the user agent can verify that the data it loads from that URL matches that expected hash before executing the JavaScript it contains. This integrity verification significantly reduces the risk that an attacker can substitute malicious content.

This example can be communicated to a user agent by adding the hash to a script element, like so:

<script src="https://example.com/example-framework.js"
        integrity="sha384-Li9vy3DqF8tnTXuiaAJuML3ky+er10rcgNR/VqsVpcw+ThHmYcwiB1pbOxEbzJr7"
        crossorigin="anonymous"></script>

Scripts, of course, are not the only response type which would benefit from integrity validation. The scheme specified here also applies to link and future versions of this specification are likely to expand this coverage.

1.1. Goals

1. Compromise of a third-party service should not automatically mean compromise of every site which includes its scripts. Content authors will have a mechanism by which they can specify expectations for content they load, meaning for example that they could load a specific script, and not any script that happens to have a particular URL.

2. The verification mechanism should have error-reporting functionality which would inform the author that an invalid response was received.

1.2. Use Cases/Examples

1.2.1. Resource Integrity

• An author wishes to use a content delivery network to improve performance for globally-distributed users. It is important, however, to ensure that the CDN’s servers deliver only the code the author expects them to deliver. To mitigate the risk that a CDN compromise (or unexpectedly malicious behavior) would change that site in unfortunate ways, the following integrity metadata is added to the link element included on the page:

  <link rel="stylesheet" href="https://site53.example.net/style.css"
        integrity="sha384-+/M6kredJcxdsqkczBUjMLvqyHb1K/JThDXWsBVxMEeZHEaMKEOEct339VItX1zB"
        crossorigin="anonymous">
  

• An author wants to include JavaScript provided by a third-party analytics service. To ensure that only the code that has been carefully reviewed is executed, the author generates integrity metadata for the script, and adds it to the script element:

  <script src="https://analytics-r-us.example.com/v1.0/include.js"
          integrity="sha384-MBO5IDfYaE6c6Aao94oZrIOiC6CGiSN2n4QUbHNPhzk5Xhm0djZLQqTpL0HzTUxk"
          crossorigin="anonymous"></script>
  

• A user agent wishes to ensure that JavaScript code running in high-privilege HTML contexts (for example, a browser’s New Tab page) aren’t manipulated before display. Integrity metadata mitigates the risk that altered JavaScript will run in these pages' high-privilege contexts.

2. Key Concepts and Terminology

This section defines several terms used throughout the document.

The term digest refers to the base64 encoded result of executing a cryptographic hash function on an arbitrary block of data.

The terms origin and same origin are defined in HTML. [HTML]

A base64 encoding is defined in Section 4 of RFC 4648. [RFC4648]

The SHA-256, SHA-384, and SHA-512 are part of the SHA-2 set of cryptographic hash functions defined by the NIST. [SHA2]

The valid SRI hash algorithm token set is the ordered set « "sha256", "sha384", "sha512" » (corresponding to SHA-256, SHA-384, and SHA-512 respectively). The ordering of this set is meaningful, with stronger algorithms appearing later in the set. See § 3.2.2 Priority and § 3.3.3 Get the strongest metadata from set for additional information.

A string is a valid SRI hash algorithm token if its ASCII lowercase is contained in the valid SRI hash algorithm token set.

2.1. Grammatical Concepts

The Augmented Backus-Naur Form (ABNF) notation used in this document is specified in RFC5234. [ABNF]

Appendix B.1 of [ABNF] defines the VCHAR (printing characters) and WSP (whitespace) rules.

Content Security Policy defines the base64-value and hash-algorithm rules. [CSP]

3. Framework

The integrity verification mechanism specified here boils down to the process of generating a sufficiently strong cryptographic digest for a resource, and transmitting that digest to a user agent so that it may be used to verify the response.

3.1. Integrity metadata

To verify the integrity of a response, a user agent requires integrity metadata as part of the request. This metadata consists of the following pieces of information:

• cryptographic hash function ("alg")

• digest ("val")

• options ("opt")

The hash function and digest MUST be provided in order to validate a response’s integrity.

Note: At the moment, no options are defined. However, future versions of the spec may define options, such as MIME types [MIME-TYPES].

This metadata MUST be encoded in the same format as the hash-source (without the single quotes) in section 4.2 of the Content Security Policy Level 2 specification.

For example, given a script resource containing only the string alert('Hello, world.');, an author might choose SHA-384 as a hash function. H8BRh8j48O9oYatfu5AZzq6A9RINhZO5H16dQZngK7T62em8MUt1FLm52t+eX6xO is the base64 encoded digest that results. This can be encoded as follows:

sha384-H8BRh8j48O9oYatfu5AZzq6A9RINhZO5H16dQZngK7T62em8MUt1FLm52t+eX6xO

echo -n "alert('Hello, world.');" | openssl dgst -sha384 -binary | openssl base64 -A

3.2. Cryptographic hash functions

Conformant user agents MUST support the SHA-256, SHA-384, and SHA-512 cryptographic hash functions for use as part of a request’s integrity metadata and MAY support additional hash functions defined in future iterations of this document.

NOTE: The algorithms supported in this document are (currently!) believed to be resistent to second-preimage and collision attacks. Future additions/removals from the set of supported algorithms would be well-advised to apply similar standard. See § 5.2 Hash collision attacks.

3.2.1. Agility

Multiple sets of integrity metadata may be associated with a single resource in order to provide agility in the face of future cryptographic discoveries. For example, the resource described in the previous section may be described by either of the following hash expressions:

sha384-H8BRh8j48O9oYatfu5AZzq6A9RINhZO5H16dQZngK7T62em8MUt1FLm52t+eX6xO
sha512-Q2bFTOhEALkN8hOms2FKTDLy7eugP2zFZ1T8LCvX42Fp3WoNr3bjZSAHeOsHrbV1Fu9/A0EzCinRE7Af1ofPrw==

Authors may choose to specify both, for example:

<script src="hello_world.js"
   integrity="sha384-H8BRh8j48O9oYatfu5AZzq6A9RINhZO5H16dQZngK7T62em8MUt1FLm52t+eX6xO
              sha512-Q2bFTOhEALkN8hOms2FKTDLy7eugP2zFZ1T8LCvX42Fp3WoNr3bjZSAHeOsHrbV1Fu9/A0EzCinRE7Af1ofPrw=="
   crossorigin="anonymous"></script>

In this case, the user agent will choose the strongest hash function in the list, and use that metadata to validate the response (as described below in the § 3.3.2 Parse metadata and § 3.3.3 Get the strongest metadata from set algorithms).

When a hash function is determined to be insecure, user agents SHOULD deprecate and eventually remove support for integrity validation using the insecure hash function. User agents MAY check the validity of responses using a digest based on a deprecated function.

To allow authors to switch to stronger hash functions without being held back by older user agents, validation using unsupported hash functions acts like no integrity value was provided (see the § 3.3.4 Do bytes match metadataList? algorithm below). Authors are encouraged to use strong hash functions, and to begin migrating to stronger hash functions as they become available.

3.2.2. Priority

The prioritization of hash algorithms is specified via the ordering of their respective tokens in the valid SRI hash algorithm token set. Algorithms appearing earlier in that set are weaker than algorithms appearing later in that set.

As currently specified, SHA-256 is weaker than SHA-384, which is in turn weaker than SHA-512. No other hashing algorithms are currently supported by this specification.

3.3. Response verification algorithms

3.3.1. Apply algorithm to bytes

1. Let result be the result of applying algorithm to bytes.

2. Return the result of base64 encoding result.

3.3.2. Parse metadata

This algorithm accepts a string, and returns a set of hash expressions whose hash functions are understood by the user agent.

1. Let result be the empty set.

2. For each item returned by splitting metadata on spaces:

   1. Let expression-and-options be the result of splitting item on U+003F (?).

   2. Let algorithm-expression be expression-and-options[0].

   3. Let base64-value be the empty string.

   4. Let algorithm-and-value be the result of splitting algorithm-expression on U+002D (-).

   5. Let algorithm be algorithm-and-value[0].

   6. If algorithm-and-value[1] exists, set base64-value to algorithm-and-value[1].

   7. If algorithm is not a valid SRI hash algorithm token, then continue.

   8. Let metadata be the ordered map «["alg" → algorithm, "val" → base64-value]».

      Note: Since no options are defined (see the § 3.1 Integrity metadata), a corresponding entry is not set in metadata. If options are defined in a future version, expression-and-options[1] can be utilized as options.

   9. Append metadata to result.

3. Return result.

3.3.3. Get the strongest metadata from set

1. Let result be the empty set and strongest be the empty string.

2. For each item in set:

   1. Assert: item["alg"] is a valid SRI hash algorithm token.

   2. If result is the empty set, then:

      1. Append item to result.

      2. Set strongest to item.

      3. Continue.

   3. Let currentAlgorithm be strongest["alg"], and currentAlgorithmIndex be the index of currentAlgorithm in the valid SRI hash algorithm token set.

   4. Let newAlgorithm be the item["alg"], and newAlgorithmIndex be the index of newAlgorithm in the valid SRI hash algorithm token set.

   5. If newAlgorithmIndex is less than currentAlgorithmIndex, continue.

   6. Otherwise, if newAlgorithmIndex is greater than currentAlgorithmIndex:

      1. Set strongest to item.

      2. Set result to « item ».

   7. Otherwise, newAlgorithmIndex and currentAlgorithmIndex are the same value. Append item to result.

3. Return result.

3.3.4. Do bytes match metadataList?

1. Let parsedMetadata be the result of parsing metadataList.

2. If parsedMetadata is empty set, return true.

3. Let metadata be the result of getting the strongest metadata from parsedMetadata.

4. For each item in metadata:

   1. Let algorithm be the item["alg"].

   2. Let expectedValue be the item["val"].

   3. Let actualValue be the result of applying algorithm to bytes .

   4. If actualValue is a case-sensitive match for expectedValue, return true.

5. Return false.

This algorithm allows the user agent to accept multiple, valid strong hash functions. For example, a developer might write a script element such as:

<script src="https://example.com/example-framework.js"
        integrity="sha384-Li9vy3DqF8tnTXuiaAJuML3ky+er10rcgNR/VqsVpcw+ThHmYcwiB1pbOxEbzJr7
                   sha384-+/M6kredJcxdsqkczBUjMLvqyHb1K/JThDXWsBVxMEeZHEaMKEOEct339VItX1zB"
        crossorigin="anonymous"></script>

which would allow the user agent to accept two different content payloads, one of which matches the first SHA-384 hash value and the other matches the second SHA-384 hash value.

Note: User agents may allow users to modify the result of this algorithm via user preferences, bookmarklets, third-party additions to the user agent, and other such mechanisms. For example, redirects generated by an extension like HTTPS Everywhere could load and execute correctly, even if the HTTPS version of a resource differs from the HTTP version.

Note: Subresource Integrity requires CORS and it is a logical error to attempt to use it without CORS. User agents are encouraged to report a warning message to the developer console to explain this failure. [Fetch]

3.4. Verification of HTML document subresources

A variety of HTML elements result in requests for resources that are to be embedded into the document, or executed in its context. To support integrity metadata for some of these elements, a new integrity attribute is added to the list of content attributes for the link and script elements. [HTML]

Note: A future revision of this specification is likely to include integrity support for all possible subresources, i.e., a, audio, embed, iframe, img, link, object, script, source, track, and video elements.

3.5. The integrity attribute

The integrity attribute represents integrity metadata for an element. The value of the attribute MUST be either the empty string, or at least one valid metadata as described by the following ABNF grammar:

integrity-metadata = *WSP hash-with-options *(1*WSP hash-with-options ) *WSP / *WSP
hash-with-options  = hash-expression *("?" option-expression)
option-expression  = *VCHAR
hash-expression    = hash-algorithm "-" base64-value

option-expressions are associated on a per hash-expression basis and are applied only to the hash-expression that immediately precedes it.

In order for user agents to remain fully forwards compatible with future options, the user agent MUST ignore all unrecognized option-expressions.

Note: Note that while the option-expression has been reserved in the syntax, no options have been defined. It is likely that a future version of the spec will define a more specific syntax for options, so it is defined here as broadly as possible.

3.6. Handling integrity violations

The user agent will refuse to render or execute responses that fail an integrity check, instead returning a network error as defined in Fetch [Fetch].

Note: On a failed integrity check, an error event is fired. Developers wishing to provide a canonical fallback resource (e.g., a resource not served from a CDN, perhaps from a secondary, trusted, but slower source) can catch this error event and provide an appropriate handler to replace the failed resource with a different one.

4. Proxies

Optimizing proxies and other intermediate servers which modify the responses MUST ensure that the digest associated with those responses stays in sync with the new content. One option is to ensure that the integrity metadata associated with resources is updated. Another would be simply to deliver only the canonical version of resources for which a page author has requested integrity verification.

To help inform intermediate servers, those serving the resources SHOULD send along with the resource a Cache-Control header with a value of no-transform.

5. Security and Privacy Considerations

This section is not normative.

5.1. Non-secure contexts remain non-secure

Integrity metadata delivered by a context that is not a Secure Context such as an HTTP page, only protects an origin against a compromise of the server where an external resources is hosted. Network attackers can alter the digest in-flight (or remove it entirely, or do absolutely anything else to the document), just as they could alter the response the hash is meant to validate. Thus, it is recommended that authors deliver integrity metadata only to a Secure Context. See also Securing the Web.

5.2. Hash collision attacks

Digests are only as strong as the hash function used to generate them. It is recommended that user agents refuse to support known-weak hashing functions and limit supported algorithms to those known to be collision resistant. Examples of hashing functions that are not recommended include MD5 and SHA-1. At the time of writing, SHA-384 is a good baseline.

Moreover, it is recommended that user agents re-evaluate their supported hash functions on a regular basis and deprecate support for those functions shown to be insecure. Over time, hash functions may be shown to be much weaker than expected and, in some cases, broken, so it is important that user agents stay aware of these developments.

5.3. Cross-origin data leakage

This specification requires integrity-protected cross-origin requests to use the CORS protocol to ensure that the resource’s content is explicitly shared with the requestor. If that requirement were omitted, attackers could violate the same-origin policy and determine whether a cross-origin resource has certain content.

Attackers would attempt to load the resource with a known digest, and watch for load failures. If the load fails, the attacker could surmise that the response didn’t match the hash and thereby gain some insight into its contents. This might reveal, for example, whether or not a user is logged into a particular service.

Moreover, attackers could brute-force specific values in an otherwise static resource. Consider a JSON response that looks like this:

{'status': 'authenticated', 'username': 'admin'}

An attacker could precompute hashes for the response with a variety of common usernames, and specify those hashes while repeatedly attempting to load the document. A successful load would confirm that the attacker has correctly guessed the username.

6. Acknowledgements

Much of the content here is inspired heavily by Gervase Markham’s Link Fingerprints concept as well as WHATWG’s Link Hashes.

A special thanks to Mike West for his invaluable contributions to the initial version of this spec. Thanks to Brad Hill, Anne van Kesteren, Jonathan Kingston, Mark Nottingham, Sergey Shekyan , Dan Veditz, Eduardo Vela, Tanvi Vyas, and Michal Zalewski for providing invaluable feedback.