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 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