fix(Headers): don't forward secure headers to 3th party

[jimmywarting]

Purpose

Avoid forwarding secure headers such as authorization, www-authenticate, cookie & cookie2 when redirecting to a untrusted site. (so it follows browses origin policy)
(matches Go behavior)

Changes

When the location headers redirects to an other hostname, strip away secure request headers meant for the original destination

Additional

Should patch v2 also (not many can update to esm)

---

• I updated ./docs/CHANGELOG.md with a link to this PR or Issue
• I updated ./docs/v3-UPGRADE-GUIDE
• I updated readme
• I added unit test(s)

---

• Fixes Potential security issue #1443

[LinusU]

What do you think about adding the . to the endsWith call?

Suggested change

	if (!a.endsWith(b)) {
	return false;
	}
	return a[a.length - b.length - 1] === '.';
	return a.endsWith(`.${b}`)

I think that this is much easier to read ☺️

[jimmywarting]

if you are making a request to fridas-blog.uk.com and it redirects to uk.com then uk.com should not know about the cookie... cuz the cookie can be tied to fridas-blog.uk.com

a = referer (original request) fridas-blog.uk.com
b = destination (Location) uk.com

a.endsWith('.${b}') === true
b is not the same host or a subdomain of a...

so i guess it can't work...

[jimmywarting]

if you are only after something that is sorter and dose the same thing:

export const isDomainOrSubdomain = (a, b) => {
  a = new URL(a).hostname
  b = new URL(b).hostname
  return a === b || (
    a[a.length - b.length - 1] === '.' && a.endsWith(b)
  )
}

[LinusU]

I still don't understand how the two checks are different? As far as I can tell, unless I'm missing something, the two functions below behave the exact same way?

function one (a, b) {
  return a === b || (
    a[a.length - b.length - 1] === '.' && a.endsWith(b)
  )
}

function two (a, b) {
  return a === b || a.endsWith(`.${b}`)
}

---

The case you are mentioning seems to be handled wrong in the committed code then?

> isDomainOrSubdomain('http://uk.com', 'http://fridas-blog.uk.com')
true

(extra ping @jimmywarting since this is already merged)

[jimmywarting]

hmm, double checked again and you are right... they are equally the same, But there is no mistake in the merged code...

I was blinded by the order of arguments passed down to isDomainOrSubdomain and mixing up whats gets passed down to the function and in which order. i was just so blinded by how Go and follow-redirects also did it using the dot index

But at least there is no rush to change it now... can change it later

[jimmywarting]

a & b was a stupid variable name...

[LinusU]

Nice 👍

Submitted PR to make the code easier to understand here: #1455

[gr2m]

Looks good to me. Feel free to refactor the isDomainOrSubdomain method to make it easier to read, but it's good as is from my perspective.

[gr2m]

Should patch v2 also (not many can update to esm)

Will you do that yourself or shall we do a follow up issue so we don't forget?

[bitinn]

Thx for everyone to get to this, this issue can finally be closed :)

But do note the drawback:

#274

[jimmywarting]

I can fix V2 also

[glasser]

Should this use the public suffix list instead? https://publicsuffix.org/

[jimmywarting]

I don't think that would be necessary.

[glasser]

The names of the variables here don't see to match the values passed in?

if (!isDomainOrSubdomain(request.url, locationURL)) {

Surely locationURL is the destination and request.url is the original?

[jimmywarting]

I think you are right. it works as intended and how it is suppose to protect you, But the variable names are mixed up

[jimmywarting]

No rush to get fixed, can create a issue about it if needed

[kentonv]

This change may match Go behavior but it differs from browser behavior and deviates from the fetch API spec.

In Chrome, if I do:

fetch(url, {headers: {"Authorization": "foo"}})

and this results in a redirect to a third party, Chrome will forward the Authorization header to the third party, assuming that the third party's CORS policy allows it. Note that the CORS policy of the original URL is irrelevant -- the third party controls the policy here. So, in an attack scenario, the attacker controls the policy, so CORS is not a defense.

AFAICT, Chrome's behavior conforms to the Fetch spec. The Fetch spec only says these headers should be removed on a redirect if they were implicitly added by the user-agent in the first place -- i.e. not when they were explicitly passed to fetch() in the request.headers attribute. Since node-fetch presumably never adds these headers implicitly, it seems like this doesn't apply.

I am concerned that this change could actually break legitimate use cases. Imagine api.example.com is a REST API that requires authorization via an Authorization header. Now imagine example.com decides to rebrand as example.org, using 3xx redirects to redirect all requests to the new domain. API clients coded against the old domain will break.

[jimmywarting]

That may be legit cases...

I have also wonder about the usage of Request.credentials if we somehow could use this in any way. The option same-origin could technically match what we have made in this change and use that as the default. and if credentials is include, then we could allow insecure headers to be forwarded?

Xhr have withCredentials
that has similar options

[kentonv]

Request.credentials seems to instruct the browser on whether or not it should automatically add cookies and HTTP basic auth credentials to the request. If the application explicitly specifies an Authorization header, though, I don't think Requset.credentials is meant to have any effect on that. In Chrome, at least, an explicitly-specified Authorization header seems to be redirected even when Request.credentials is same-origin (the default).

[glasser]

@kentonv (Hi!) Just to check, your concern is specifically about Authorization, right? It does seem reasonable to me that cookies should not be sent to a different origin.

[kentonv]

Hi @glasser! Long time no see.

TBH, I'm not really sure what this should mean for Cookie. Technically, the fetch() spec does not allow the application to specify this header at all, so for a "conformant" implementation the question is moot. But, of course, that part of the fetch spec really only makes sense for browsers.

What makes sense on the server side? I don't really know because I don't know under what circumstances a server would ever send a Cookie header. I'd expect Node apps are mainly using fetch() to call APIs, and APIs rarely if ever use Cookie.

For Cloudflare Workers (of which I'm the lead engineer), our fetch() implementation does have to allow the Cookie header for the proxying use case: many workers receive a request from a client, rewrite it in some way, and then pass it on to an origin server using fetch(). The Cookie header must come along for that ride -- and it's important that it do so even if the hostname changed in the process. But, in CF Workers, proxied requests always have redirect = "manual", because a redirect response should always be returned back to the original client rather than handled within the worker. So, what happens to Cookie on a redirect is moot for the common case.

A Worker can also make a stand-alone fetch() that is not proxying any existing request, and in that case the default redirect = "follow" applies. But this is again typically used for calling APIs, and I don't know why Cookie would be used at all, so it's hard to say what the right behavior is when redirecting.

FWIW, at present Cloudflare Workers has no special logic whatsoever for Cookie or Authorization -- these are just plain old headers to us.

Of course, for the API use case, it's very common for other headers to contain sensitive information, too. Some APIs use non-standard headers like X-Auth-Key for their secrets. But also if the API uses any special headers at all, who knows what is in them? So it feels weird to me to even try to protect this use case from leaking secrets on a redirect... it seems like it's really the application's responsibility to use manual redirect logic if they don't want the whole request to be forwarded.

So in the absence of more information about real-world use cases, my inclination is to say that there's no reason to treat Cookie specially here. But I can definitely see where it is debatable.

[jimmywarting]

Do you know what alt-sve will do with cookies and other headers if it decided to try another url the next time?

[jroitgrund]

What is the recommended way of reproducing the old behavior now?

[pkit]

@jroitgrund I suppose the only way is to set redirect: "manual" and then manually follow the redirects with another fetch() call.
It's a bummer that this issue even exists. I will see if it's easy to patch locally and will post it.

[pkit]

@jroitgrund here's the patch changes the behavior to the old one.
https://gist.github.com/pkit/a65456cd29505c1d12346f8d57a45c2e