Position on User-Agent Client Hints

# Yoav Weiss (2 days ago)

Hey WebKit folks,

I'd like to ask for your official position on the User Agent Client Hints wicg.github.io/ua-client-hints specification.

+Maciej Stachowiak <mjs at apple.com> has been providing great feedback on the

proposal, as well as to the underlying Client Hints Infrastructure wicg.github.io/client-hints-infrastructure (in its former

iterations), which helped shape those proposals, but that obviously doesn't count as endorsement.

There's an intent to ship for the feature groups.google.com/a/chromium.org/d/msg/blink-dev/A4wxFpvqUfA/g7iccl9ICgAJ

over at blink-dev, and your position would be appreciated as input into that process.

Thanks :) Yoav

Contact us to advertise here
# Michael Catanzaro (a day ago)

My personal $0.02: I'm mildly supportive of this spec. It's certainly an improvement on existing HTTP user agent headers. I appreciate that you worked to incorporate feedback into the spec and considered the concerns of small browsers.

Is it going to solve all the problems caused by user agent headers? No. If WebKit implements the spec, we're surely going to eventually need a quirks list for user agent client hints to decide which websites to lie to, just like we already have quirks for the user agent header. And as long as Chrome sends a user agent header that includes the string "Chrome", it's unlikely we'll be able to get rid of the existing quirks list. But I think client hints will probably reduce the amount of websites that accidentally break WebKit, by replacing wild west UA header parsing with well-defined APIs, and adding some GREASE for good measure. The promise of freezing Chrome's UA header sounds nice, as it makes quirks easier to maintain. And being able to ration entropy by revealing details about the platform on an active rather than passive basis is quite appealing.

The spec attracted some misplaced concern about negative impact to small browsers, which I've rebutted in [1]. I'm not quite so enthusiastic about this spec as I was initially, especially after I was convinced that the GREASE is never going to be enough to remove our quirks list, but it's certainly not going to hurt small browsers.

This spec has received some pretty harsh criticism from the user tracking industry (some call it the "ad industry"). Not historically a friend of WebKit, so sounds good to me. ;)

One concern I haven't mentioned elsewhere is that frozen UA header might encourage deeper levels of fingerprinting than are currently used, e.g. for ad fraud prevention. caddy has started blocking WebKitGTK users based on TLS handshake fingerprint (yes, really!) [1]. If techniques like that take off as a result of this, that could potentially backfire on us quite badly. But websites could choose to do such things today anyway, client hints or no, and if so, the solution will be for us to just try even harder to look more like Chrome.

Seems like a net positive overall. I don't work for Apple and can't say whether it might be implemented by WebKit.

Michael

[1] w3ctag/design-reviews#467 [2] mitm.watch

# Haelwenn (lanodan) Monnier (a day ago)

[2020-05-07 18:22:10-0500] Michael Catanzaro:

My personal $0.02: I'm mildly supportive of this spec. It's certainly an improvement on existing HTTP user agent headers. I appreciate that you worked to incorporate feedback into the spec and considered the concerns of small browsers.

Is it going to solve all the problems caused by user agent headers? No. If WebKit implements the spec, we're surely going to eventually need a quirks list for user agent client hints to decide which websites to lie to, just like we already have quirks for the user agent header. And as long as Chrome sends a user agent header that includes the string "Chrome", it's unlikely we'll be able to get rid of the existing quirks list. But I think client hints will probably reduce the amount of websites that accidentally break WebKit, by replacing wild west UA header parsing with well-defined APIs, and adding some GREASE for good measure. The promise of freezing Chrome's UA header sounds nice, as it makes quirks easier to maintain. And being able to ration entropy by revealing details about the platform on an active rather than passive basis is quite appealing.

The spec attracted some misplaced concern about negative impact to small browsers, which I've rebutted in [1]. I'm not quite so enthusiastic about this spec as I was initially, especially after I was convinced that the GREASE is never going to be enough to remove our quirks list, but it's certainly not going to hurt small browsers.

This spec has received some pretty harsh criticism from the user tracking industry (some call it the "ad industry"). Not historically a friend of WebKit, so sounds good to me. ;)

One concern I haven't mentioned elsewhere is that frozen UA header might encourage deeper levels of fingerprinting than are currently used, e.g. for ad fraud prevention. caddy has started blocking WebKitGTK users based on TLS handshake fingerprint (yes, really!) [2]. If techniques like that take off as a result of this, that could potentially backfire on us quite badly. But websites could choose to do such things today anyway, client hints or no, and if so, the solution will be for us to just try even harder to look more like Chrome.

Seems like a net positive overall. I don't work for Apple and can't say whether it might be implemented by WebKit.

Michael

[1] w3ctag/design-reviews#467 [2] mitm.watch

This particular thing is bullshit, it's not even using the TLS handshake. Copying the curl options is enough and what made me notice it that glib-networking with OpenSSL (non-default) has the message.

Screenshot showing a simple curl vs. a curl with webkit-gtk's options: queer.hacktivis.me/media/5d9122fd-64c4-43de-ba02-776b162c106e/screen.png

Most aggressive fingerprinting done currently is done with JavaScript or CSS, having yet another few lines to get the Client Hints isn't going to change much things for them. Sure, it makes it a bit harder for some small advertisers that wouldn't do much fingerprinting but do not forget that Google is probably the biggest one and with a very large toolset.

# Maciej Stachowiak (a day ago)

I would consider myself mildly positive as to the direction, but that’s my personal view for the moment, absent consultation with my colleagues. I will solicit more viewpoints.

I particularly appreciate the responsiveness to feedback and that Yoav in particular has been willing to iterate.

I think there’s a number of things in the spec that should be cleaned up before an implementation ships enabled by default, specifically around interop, privacy, and protection against UA lockouts. I know there are PRs in flight for some of these issues. I think it would be good to get more of the open issues to resolution before actually shipping this.

# Maciej Stachowiak (a day ago)

Accidentally removed Yoav from Cc and I’m not sure if he is on this list.

# Alex Christensen (18 hours ago)

Last I recall talking about this, we did not oppose to client hits header fields in general, just some specific ones that expose information we do not wish to expose to reduce fingerprinting information. For example, I think we do oppose adding Device-Memory because that currently cannot be queried through WebKit any other way, but I don’t think we oppose adding Viewport-Width which is trivial to query with 100% accuracy once you have JavaScript on the client. I think Downlink and RTT were in a grey area because they can indeed be measured other ways, but we don’t want to possibly increase the fingerprinting information by providing values that can be used for more effective client fingerprinting, such as if we were to send the exact same value to multiple servers.

I don’t think this is an official position, just what I recall from TPAC discussions in Lyon.

# Yoav Weiss (11 hours ago)

On Fri, May 8, 2020 at 6:01 PM Alex Christensen <achristensen at apple.com>

wrote:

Last I recall talking about this, we did not oppose to client hits header fields in general, just some specific ones that expose information we do not wish to expose to reduce fingerprinting information. For example, I think we do oppose adding Device-Memory because that currently cannot be queried through WebKit any other way, but I don’t think we oppose adding Viewport-Width which is trivial to query with 100% accuracy once you have JavaScript on the client. I think Downlink and RTT were in a grey area because they can indeed be measured other ways, but we don’t want to possibly increase the fingerprinting information by providing values that can be used for more effective client fingerprinting, such as if we were to send the exact same value to multiple servers.

I don’t think this is an official position, just what I recall from TPAC discussions in Lyon.

Thanks! Just noting that this request is about User-Agent Client Hints in particular.

On May 8, 2020, at 12:14 AM, Maciej Stachowiak <mjs at apple.com> wrote:

Accidentally removed Yoav from Cc and I’m not sure if he is on this list.

I'm still subscribed :D

>

On May 8, 2020, at 12:04 AM, Maciej Stachowiak <mjs at apple.com> wrote:

I would consider myself mildly positive as to the direction, but that’s my personal view for the moment, absent consultation with my colleagues. I will solicit more viewpoints.

I particularly appreciate the responsiveness to feedback and that Yoav in particular has been willing to iterate.

Thank you! :)

>

I think there’s a number of things in the spec that should be cleaned up before an implementation ships enabled by default, specifically around interop, privacy, and protection against UA lockouts. I know there are PRs in flight for some of these issues. I think it would be good to get more of the open issues to resolution before actually shipping this.

Regarding the discussion around architecture, my thinking is that we could either reach a conclusion there before shipping, or ship with that value currently empty and "fill it" once we reach agreement on what it should be.

Regarding the GREASE discussion, I don't think it's blocking, as Chromium's initial implementation will include GREASE: The UA will be a list of values, will include a non-value that would encourage standard Structured Header parsing, and will have the list order randomized (but remain stable for the lifetime of the major version, the avoid caching churn). With that said, your arguments about making this a MUST make sense, and I'll send a PR to that effect.

The other issues seem to be either around future enhancement (which we could add later), or general discussions that don't seem blocking.

Want more features?

Request early access to our private beta of readable email premium.