Prefer Secure Origins For Powerful New Features

We (Chrome Security) originally sent this out to various browser development mailing lists. Here is the canonical location for the original proposal. See this link for the current public draft spec.

This is a living document — as we learn more, we'll probably need to change this page.

Proposal

The Chrome Security team and I propose that, for new and particularly powerful web platform features, browser vendors tend to prefer to make the the feature available only to secure origins by default.

Definitions

“Particularly powerful” would mean things like: features that handle personally-identifiable information, features that handle high-value information like credentials or payment instruments, features that provide the origin with control over the UA's trustworthy/native UI, access to sensors on the user's device, or generally any feature that we would provide a user-settable permission or privilege to. Please discuss!

“Particularly powerful” would not mean things like: new rendering and layout features, CSS selectors, innocuous JavaScript APIs like showModalDialog, or the like. I expect that the majority of new work in HTML5 fits in this category. Please discuss!

“Secure origins” are origins that match at least one of the following (scheme, host, port) patterns:

• (https, *, *)
• (wss, *, *)
• (*, localhost, *)
• (*, 127/8, *)
• (*, ::1/128, *)
• (file, *, —)
• (chrome-extension, *, —)

This list may be incomplete, and may need to be changed. Please discuss! A bug to define “secure transport” in Blink/Chromium: https://code.google.com/p/chromium/issues/detail?id=362214

For Example

For example, Chrome is going to make Service Workers available only to secure origins, because it provides the origin with a new, higher degree of control over a user's interactions with the origin over an extended period of time, and because it gives the origin some control over the user's device as a background task.

Consider the damage that could occur if a user downloaded a service worker script that had been tampered with because they got it over a MITM’d or spoofed cafe wifi connection. What should have been a nice offline document editor could be turned into a long-lived spambot, or maybe even a surveillance bot. If the script can only run when delivered via authenticated, integrity-protected transport like HTTPS, that particular risk is significantly mitigated.

Background

Legacy platforms/operating systems have a 1-part principal: the user. When a user logs in, they run programs that run with the full privilege of the user: all of a user’s programs can do anything the user can do on all their data and with all their resources. This has become a source of trouble since the rise of mobile code from many different origins. It has become less and less acceptable for a user’s (e.g.) word processor to (e.g.) read the user’s private SSH keys.

Modern platforms have a 2-part security principal: the user, and the origin of the code. Examples of such modern platforms include (to varying degrees) the web, Android, and iOS. In these systems, code from one origin has (or, should have) access only to the resources it creates and which are explicitly given to it.

For example, the Gmail app on Android has access only to the user’s Gmail and the system capabilities necessary to read and write that email. Without an explicit grant, it does not have access to resources that other apps (e.g. Twitter) create. It also does not have access to system capabilities unrelated to email. Nor does it have access to the email of another user on the same computer.

In systems with 2-part principals, it is crucial to strongly authenticate both parts of the principal, not just one part. (Otherwise, the system essentially degrades into a 1-part principal system.) This is why, for example, both Android and iOS require that every vendor (i.e. origin) cryptographically sign its code. That way, when a user chooses to install Twitter and to give Twitter powerful permissions (such as access to the device’s camera), they can be sure that they are granting such capability only to the Twitter code, and not to just any code.

By contrast, the web has historically made origin authentication optional. On the web, origins are defined as having 3 parts: (scheme, host, port), e.g. (HTTP, example.com, 80) or (HTTPS, mail.google.com, 443). Many origins use unauthenticated schemes like HTTP, WS, or even FTP.

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.

Thank You For Reading This Far!

We welcome discussion, critique, and cool new features!