I have simplified it to

== <center><img src=ignored onerror=alert(1)><img src=triggers onerror=alert(document.domain)> ==

You need a header with a center tag, and the onerror of the first img doesn't run, but the second does.

Only executed on mobile

The first img doesn't really need any parameters:

==  <center><img><img src=zxcv onerror=throw(document.domain)> ==

This html is copied verbatim (but for the center) inside title of the href:

<a href="/w/index.php?title=Wikipedia:Sandbox/2020-09-07&amp;action=edit&amp;section=1" title="Edit section: <img><img src=zxcv onerror=throw(document.domain)>"

Thanks @Platonides. Changing your example to

== <center><img src=ignored onerror=alert(1)><img src=triggers onerror=console.trace() ==

includes methods called initialize, _postInitialize, and render, which point to View#render in MobileFrontend/src/mobile.startup/View.js (https://gerrit.wikimedia.org/g/mediawiki/extensions/MobileFrontend/+/d4ae221a534a31d498aa3c8d86c796cf675c3e71/src/mobile.startup/View.js#252).

If the page is protected (thus no edit section link), the XSS doesn't fire

This also happens (in Mobile) when forcing a different skin, such as monobook or vector

It looks like this line 56 from PageGateway.js [1] is at least somewhat implicated in this:

section.line = section.line.replace( /<\/?a\b[^>]*>/g, '' );

Line 212 first gets the html from the .mw-headline selector [2] and then I think the replace call above is stripping tags and is making what was once inert HTML into an XSS attack vector

[1] https://github.com/wikimedia/mediawiki-extensions-MobileFrontend/blob/master/src/mobile.startup/PageGateway.js#L56
[2] https://github.com/wikimedia/mediawiki-extensions-MobileFrontend/blob/d4ae221a534a31d498aa3c8d86c796cf675c3e71/src/mobile.startup/PageGateway.js#L212

To confirm what @nray says above:

<center>&lt;img src=ignored onerror=alert(1)&gt;&lt;img src=triggers onerror=console.trace();&gt;<span class="mw-editsection"><a href="/w/index.php?title=Talk:T262213&amp;action=edit&amp;section=1" title="Edit section: <img src=ignored onerror=alert(1)><img src=triggers onerror=console.trace();>" data-section="1" class="mw-ui-icon mw-ui-icon-element mw-ui-icon-wikimedia-edit-base20 edit-page mw-ui-icon-flush-right">Edit</a></span></center>

gets converted to

<center>&lt;img src=ignored onerror=alert(1)&gt;&lt;img src=triggers onerror=console.trace();&gt;<span class="mw-editsection"><img src=triggers onerror=console.trace();>" data-section="1" class="mw-ui-icon mw-ui-icon-element mw-ui-icon-wikimedia-edit-base20 edit-page mw-ui-icon-flush-right">Edit</span></center>

probably fixed by changing to

section.line = section.line.replace(  /<\/?a\b("[^"]*"|[^>])*>/g, '' );

Or if we also want to take into account parameters using single quotes (which don't seem to be used)

section.line = section.line.replace(  /<\/?a\b("[^"]*"|'[^']*'|[^>])*>/g, '' );

Brief update: There's a patch inbound from @nray. We've got a good handle on the issue.

In T262213#6441220, @Platonides wrote:

section.line = section.line.replace(  /<\/?a\b("[^"]*"|'[^']*'|[^>])*>/g, '' );

After 2 hours fighting with mediawiki/MF/webpack: yes, it seems to work

In T262213#6441309, @phuedx wrote:

Brief update: There's a patch inbound from @nray. We've got a good handle on the issue.

Thanks!

Do you need any help with deploying it etc? I'm on slack and irc. I am going to sort some food, but not far away from my keyboard.

The basic fix I tried

plus the full change including the autogenerated files

Thank you @Platonides for the patches! I think those patches fix the immediate issue, however after discussing this with @phuedx today, we'd both like to try removing the lines relating to the regex altogether as their continued usage is questionable at best. The small patch below applies this removal plus the required webpack build artifacts

@Reedy could you help with the deploy? I do not have deploy rights

Actually removing the regex seems preferable, indeed.
However, I think this may produce links inside links, which the previous code was trying to avoid?

Testing it.

• It does fix the vulnerability
• If there is a header with links (e.g. == [[page2]] [[page3]] [[page4]] ==) you need to click it _outside_ the links to expand the section. Will that be confusing or people will manage fine? A question for usability team, I guess.

Seems to be the same that it did before, so I don't see what was that code supposed to be doing, then.

@Platonides it's important to note that the code in question doesn't seem to affect the rendering of the section headings in the DOM regardless if the regex is there or not - that's server rendered. If you checkout master, the header with links still exists with your example. As far as I know, that's considered a feature (although it might be good to review that later on).

The closest thing we could find to a reason for the current code is found at commit sha 78f85803f64ae3ecedbecb38473ee70606fca5c9 which suggests it was originally intended to fix a previous XSS issue (ironically), but I'm not able to find a reason for its existence today.

I thought it was removing links from headers, but it seems it was not doing anything ¯\_(ツ)_/¯ (other than adding a security vulnerability).

Good to go then?

@nray Yeah, I can deploy. Can you come on IRC for the deploy for some extra testing?

My +2 to nray patch

@Reedy yes, I'm on (nray is my nick)

We should get a CVE for this extension vulnerability. This code has been here since 2014, and was added itself to avoid a XSS, so basically (assuming it wasn't safe before and something changed) everyone with MobileFrontend installed would be affected.

Below adds patch on top of the origin/wmf/1.36.0-wmf.6 branch per @Reedy 's request

In T262213#6441414, @Platonides wrote:

We should get a CVE for this extension vulnerability. This code has been here since 2014, and was added itself to avoid a XSS, so basically (assuming it wasn't safe before and something changed) everyone with MobileFrontend installed would be affected.

Yeah, that's why I tagged it against T256342: Write and send supplementary release announcement for extensions and skins with security patches (1.31.9/1.34.3/1.35.0). Between that and T256341: Obtain CVEs for 1.31.9/1.34.3/1.35.0 security releases, Scott or I would usually do it as part of the process

Just putting this as a subtask beneath T257976: 1.36.0-wmf.8 deployment blockers incase of security patch issues. Not actually blocker, but keeps the visibility incase of any issues

Releng: Feel free to just remove it if all is good :)

Here's a brief writeup of the problem and why the fix makes sense:

1. The server was/is sending properly escaped content to the client
2. mobile.init/mobile.init.js calls Skin::getSingleton
3. Skin::getSingleton calls currentPage::loadCurrentPage
4. loadCurrentPage calls PageGateway::getSectionsFromHTML

PageGateway::getSectionsFromHTML finds all headings in the page content, extracts their HTML content (via $( el ).html()) and runs the following on each:

// Elsewhere
section.line = $( el ).html();

section.line = section.line.replace(  /<\/?a\b("[^"]*"|[^>])*>/g, '' )

5. The array returned by PageGateway::getSectionsFromHTML is used to construct an array of Section objects
6. The Section constructor function calls the View constructor function
7. The View constructor function:
   1. Calls $.parseHTML( 'div' ) with the current document context (the default behaviour is to use a new document – see https://api.jquery.com/jQuery.parseHTML/ for detail)
   2. Renders an inline Mustache template that uses the value of section.line unescaped
   3. Sets the inner HTML of the element created in A to the rendered template

As far as @nray and I can tell, the Section objects and their underlying HTML elements are never used by Skin::getSingleton, currentPage::loadCurrentPage, or PageGateway::getSectionsFromHTML.

This vulnerability is a result of 4 and 7. In T262213#6441157, I noted that

<center>&lt;img src=ignored onerror=alert(1)&gt;&lt;img src=triggers onerror=console.trace();&gt;<span class="mw-editsection"><a href="/w/index.php?title=Talk:T262213&amp;action=edit&amp;section=1" title="Edit section: <img src=ignored onerror=alert(1)><img src=triggers onerror=console.trace();>" data-section="1" class="mw-ui-icon mw-ui-icon-element mw-ui-icon-wikimedia-edit-base20 edit-page mw-ui-icon-flush-right">Edit</a></span></center>

(the HTML content of a heading) is converted to

<center>&lt;img src=ignored onerror=alert(1)&gt;&lt;img src=triggers onerror=console.trace();&gt;<span class="mw-editsection"><img src=triggers onerror=console.trace();>" data-section="1" class="mw-ui-icon mw-ui-icon-element mw-ui-icon-wikimedia-edit-base20 edit-page mw-ui-icon-flush-right">Edit</span></center>

In 7C, the inner HTML of the newly created element in the current document is set to the above, the UA fails to fetch "/triggers", and the error event handler is run.

We could have fixed this by:

• Improving the regular expression in 4 (see T262213#6441220)
• Override View::parseHTML in Section and don't pass $.parseHTML the current document as context
• Update View::parseHTML to the above for all View implementations
• Update Section's template to escape the value of section.line

However, @nray and I ultimately decided to remove the regular expression in 4 because, as I've noted above, the Section objects and their underlying HTML elements are never used.

Here's a list of follow-up work that @nray and I identified:

• Update SkinMinerva::doEditSectionLink to render the editsectionhint message the same way that Skin::doEditSectionLink does
• Investigate and document whether it's necessary to model the sections of the page
  • If so, make the Section class a Plain Ol' JavaScript Object – a simple, side-effect-free model not a View implementation
• Investigate whether the current behaviour of View::parseHTML is the correct behaviour

A couple more notes about the scope of the attack:

• It doesn't work when JavaScript is disabled
• It doesn't work when edit links aren't present (as @Platonides notes in T262213#6441110)
• It works on all other mobile pageviews

Thanks for the writeup.

As a plan for going forward...

Are you happy for the patch to go into gerrit? This is going to be needed "soon" as basically every patch that changes resources/dist/mobile.common.js and resources/dist/mobile.common.js.map.json is going to cause the security patch to not apply as part of the train. And as it's not just a trivial rebase (as it's a rebase + running webpack build stuff), it's more work than most patches.

As the extension isn't bundled in the tarball, and is already patched on WMF deployment, this is fairly usual practice. CVE and that disclosure can be done in the near future as part of the next security release where we generally announce these changes more widely (ala T256342).

Then it's in master and can ride the train going forward.

I would appreciate some help doing the REL1_31/REL1_34/REL1_35 backports too if possible (again because you have dev/build environments setup for MF and webpack. I think REL1_31 is pre-webpack, so it looks like it should just need the change in resources/mobile.startup/PageGateway.js, not mobile.common.js/mobile.common.js.map.json). But with the task public and patches in gerrit, this can be done easily in the open. And then merged as appropriate when CI is happy.

What about this task to being opened up and made public? Any reason we cannot do that?

Where do you want to do the followup work identified? File seperate (security as approrpiate) tasks for them?

In T262213#6443548, @Reedy wrote:

Are you happy for the patch to go into gerrit? This is going to be needed "soon" as basically every patch that changes resources/dist/mobile.common.js and resources/dist/mobile.common.js.map.json is going to cause the security patch to not apply as part of the train. And as it's not just a trivial rebase (as it's a rebase + running webpack build stuff), it's more work than most patches.

I don't see any reason why the patch can't go into Gerrit and this task made public. @dcipoletti?

I would appreciate some help doing the REL1_31/REL1_34/REL1_35 backports too if possible (again because you have dev/build environments setup for MF and webpack. I think REL1_31 is pre-webpack, so it looks like it should just need the change in resources/mobile.startup/PageGateway.js, not mobile.common.js/mobile.common.js.map.json). But with the task public and patches in gerrit, this can be done easily in the open. And then merged as appropriate when CI is happy.

I can help with this but it'll have to be tomorrow (BST).

Where do you want to do the followup work identified? File seperate (security as approrpiate) tasks for them?

I can also file the tasks as they're all work for Readers Web. Again, this'll have to be tomorrow (BST).

Yeah, no rush on my part. Definitely doesn't need doing "now".

We can certainly wait until tomorrow before putting patches into master in gerrit, opening up tasks etc :).

Branch was cut yesterday evening. The patch here (as was predicted) doesn't apply since it was generated by a build step that we don't have a good means of replicating. We can backport patches when they're available. For now continuing rollout to test wikis.

https://github.com/wikimedia/mediawiki-extensions-MobileFrontend/blob/wmf/1.36.0-wmf.6/resources/dist/mobile.common.js - 3fb4e30 (25 Jul)
https://github.com/wikimedia/mediawiki-extensions-MobileFrontend/blob/wmf/1.36.0-wmf.7/resources/dist/mobile.common.js - 3fb4e30 (25 Jul)
https://github.com/wikimedia/mediawiki-extensions-MobileFrontend/blob/wmf/1.36.0-wmf.8/resources/dist/mobile.common.js - da3edce (3 Sep)

I put the first file into the .7 folder in /srv/patches, but didn't try to apply it as obviously .7 wasn't deployed.

Speaking to Jeena on IRC... It seems, for whatever reason, the patch from .6 was rolled forward into the .8 folder... Which is obviously wrong. But I don't know how the automated copying of security patches actually works... I'd presumed it would take it from .7 as the newest version... but seemingly not? If not.. Why does .7 exist on disk etc?

reedy@deploy1001:/srv/patches$ md5sum 1.36.0-wmf.6/extensions/MobileFrontend/01-T262213.patch 
13590667c0c0274009075cd92870a24e  1.36.0-wmf.6/extensions/MobileFrontend/01-T262213.patch
reedy@deploy1001:/srv/patches$ md5sum 1.36.0-wmf.7/extensions/MobileFrontend/01-T262213.patch 
847ad3264f5bd4a15d098e88c2f18920  1.36.0-wmf.7/extensions/MobileFrontend/01-T262213.patch
reedy@deploy1001:/srv/patches$ md5sum 1.36.0-wmf.8/extensions/MobileFrontend/01-T262213.patch 
13590667c0c0274009075cd92870a24e  1.36.0-wmf.8/extensions/MobileFrontend/01-T262213.patch

Of course, things like this are evidence why these "build steps" are not helpful for security patches and wmf deployment

reedy@deploy1001:/srv/mediawiki-staging/php-1.36.0-wmf.8/extensions/MobileFrontend$ git am /srv/patches/1.36.0-wmf.7/extensions/MobileFrontend/01-T262213.patch
Applying: Remove regex section line replacement from PageGateway
reedy@deploy1001:/srv/mediawiki-staging/php-1.36.0-wmf.8/extensions/MobileFrontend$

Tidied up the patches to match reality:

reedy@deploy1001:/srv/patches$ cp 1.36.0-wmf.7/extensions/MobileFrontend/01-T262213.patch 1.36.0-wmf.8/extensions/MobileFrontend/01-T262213.patch
cp: cannot create regular file '1.36.0-wmf.8/extensions/MobileFrontend/01-T262213.patch': Permission denied
reedy@deploy1001:/srv/patches$ ls -al 1.36.0-wmf.8/extensions/MobileFrontend/01-T262213.patch
-rw-r--r-- 1 jhuneidi wikidev 8436 Sep  8 17:39 1.36.0-wmf.8/extensions/MobileFrontend/01-T262213.patch
reedy@deploy1001:/srv/patches$ ls -al 1.36.0-wmf.8/extensions/MobileFrontend/01-T262213.patch
-rw-rw-r-- 1 jhuneidi wikidev 8436 Sep  8 17:39 1.36.0-wmf.8/extensions/MobileFrontend/01-T262213.patch
reedy@deploy1001:/srv/patches$ cp 1.36.0-wmf.7/extensions/MobileFrontend/01-T262213.patch 1.36.0-wmf.8/extensions/MobileFrontend/01-T262213.patch
reedy@deploy1001:/srv/patches$ cp 1.36.0-wmf.6/extensions/MobileFrontend/01-T262213.patch 1.36.0-wmf.7/extensions/MobileFrontend/01-T262213.patch
reedy@deploy1001:/srv/patches$ git status
On branch master
Changes not staged for commit:
  (use "git add <file>..." to update what will be committed)
  (use "git checkout -- <file>..." to discard changes in working directory)

	modified:   1.36.0-wmf.7/extensions/MobileFrontend/01-T262213.patch
	modified:   1.36.0-wmf.8/extensions/MobileFrontend/01-T262213.patch

Untracked files:
  (use "git add <file>..." to include in what will be committed)

	1.36.0-wmf.8/extensions/MobileFrontend/01-T262213.patch.failed

no changes added to commit (use "git add" and/or "git commit -a")
reedy@deploy1001:/srv/patches$ rm 1.36.0-wmf.8/extensions/MobileFrontend/01-T262213.patch.failed
rm: remove write-protected regular file '1.36.0-wmf.8/extensions/MobileFrontend/01-T262213.patch.failed'? y
(failed reverse-i-search)`sha': nano T247149.^C 
reedy@deploy1001:/srv/patches$ md5sum 1.36.0-wmf.8/extensions/MobileFrontend/01-T262213.patch 
847ad3264f5bd4a15d098e88c2f18920  1.36.0-wmf.8/extensions/MobileFrontend/01-T262213.patch
reedy@deploy1001:/srv/patches$ md5sum 1.36.0-wmf.7/extensions/MobileFrontend/01-T262213.patch 
13590667c0c0274009075cd92870a24e  1.36.0-wmf.7/extensions/MobileFrontend/01-T262213.patch
reedy@deploy1001:/srv/patches$ md5sum 1.36.0-wmf.6/extensions/MobileFrontend/01-T262213.patch 
13590667c0c0274009075cd92870a24e  1.36.0-wmf.6/extensions/MobileFrontend/01-T262213.patch
reedy@deploy1001:/srv/patches$ git commit -a -m "Move T262213 .7 patch to .8, copy .6 patch to .7
> 
> Bug: T262213"
[master 5335ca0] Move T262213 .7 patch to .8, copy .6 patch to .7
 2 files changed, 0 insertions(+), 0 deletions(-)
 rename {1.36.0-wmf.8 => 1.36.0-wmf.7}/extensions/MobileFrontend/01-T262213.patch (100%)
 rename {1.36.0-wmf.7 => 1.36.0-wmf.8}/extensions/MobileFrontend/01-T262213.patch (100%)
reedy@deploy1001:/srv/patches$

Jeena just said she accidentally copied the .6 patches to .8... which explains the problems/confusion with the patches.

Maybe something we should better document this, expecially in these cases where we still branch versions (ie .7 in this case), but don't stage/deplloy it to the deployment server and as such, WMF production.

No harm done :)

Untagging as blocking T257976: 1.36.0-wmf.8 deployment blockers as that is sorted now, and as such, doesn't block the train.

Aim is to get it into master this week, so shouldn't be a blocker for T257977: 1.36.0-wmf.9 deployment blockers either

@dcipoletti has confirmed that, at least from our side, that this task can be opened up via Slack.

AIUI backports to REL1_{31,34,35} still need to be done. Is that correct?

In T262213#6446168, @phuedx wrote:

@dcipoletti has confirmed that, at least from our side, that this task can be opened up via Slack.

Cool, we can do that as patches start going up

In T262213#6446169, @phuedx wrote:

AIUI backports to REL1_{31,34,35} still need to be done. Is that correct?

Yes please! Nicks original patch should apply cleanly to master. So should be fine to go straight into gerrit

Typically, REL1_35 woudl've almost been ok with the .6 patch, but there's one commit more in 1.36 vs 1.35

Hi, just catching up on the comments here and clarifying that originally, the .6 version was copied into the .8 folder (talking this over later with thcipriani, we conjectured that this was because there was actually no .7 deploy)

Then, after talking to Reedy, when attempting to copy .7 version into the .8 folder, I mistakenly copied the .6 version again :(.

In T262213#6451366, @phuedx wrote:

I'm doing the backports now…

We're just waiting for confirmation that the original reporter wants to be credited (per https://www.mediawiki.org/wiki/Reporting_security_bugs#Crediting_Reporters).

Patches to master, REL1_{31,34,35} have been submitted and merged.