When a cached page is being restored, and the page that's being navigated away is not cacheable,
there exists a time frame during which two documents are attached to the same frame. If an attacker
finds a way to run JS during this time frame, she will be able to use one of the documents to
execute JavaScript in the context of the other one.
One possible call stack that might lead to JS execution is:
```
a child frame's unload handler
...
ContainerNode::disconnectDescendantFrames()
Document::prepareForDestruction()
FrameLoader::clear()
FrameLoader::open()
```
By the time `FrameLoader::clear` is called, child frames are usually already disconnected from the
document via
```
FrameLoader::detachChildren()
FrameLoader::setDocumentLoader()
FrameLoader::transitionToCommitted()
```
However, the attacker can initiate a new page load inside `detachChildren` to bypass
`SubframeLoadingDisabler` and create a new child frame. Note that it won't cancel the cached page
load.
The attack has a restriction that significantly limits its applicability -- a victim page should
load a (potentially sandboxed) <iframe> with attacker-controlled content, so the attacker's JS has
a chance to run inside `Document::prepareForDestruction`. This is the case, for example, for online
translators.
VERSION
WebKit revision 246194
It's unclear whether the bug is exploitable in Safari 12.1.1. The repro case seem to have an issue
with a nested `showModalDialog` call.
REPRODUCTION CASE
The test case again relies on `showModalDialog` to perform synchronous page loads. Moreover, the
code is wrapped inside a `showModalDialog` call to keep a user gesture token active throughout its
execution.
CREDIT INFORMATION
Sergei Glazunov of Google Project Zero
Proof of Concept:
https://github.com/offensive-security/exploitdb-bin-sploits/raw/master/bin-sploits/47453.zip