##How Homakov hacked GitHub and the line of code that could have prevented it
Please note: THIS ARTICLE IS NOT WRITTEN BY THE GITHUB TEAM or in any way associated with them. It's simply hosted as a Gist because the markdown formatting is excellent and far clearer than anything I could manage on my personal Tumblr at peternixey.com.
If you'd like to follow me on twitter my handle is @peternixey
@homakov’s explot on GitHub was simple and straightforward. Calling it an attack makes it sound malicious whereas the truth was that GitHub bolted its front door but left the hinges on quick release. Homakov released the hinges, walked in and shouted to anyone who would listen that they had a problem.
He was right. The Rails defaults are vulnerable and there’s no better illustration of this than when when one of the best Rails teams in the world is severely compromised.
TL;DR: How to protect your Rails application from the GitHub attack
Add the following initializer:
config/initializers/disable_mass_assignment.rb
ActiveRecord::Base.send(:attr_accessible, nil)
(this fix is not without its pitfalls - see later for things to watch for)
What the initializer does
The initalizer forces you to declare parameters that can be updated via the update_attributes
method. Rails’ default position is that any attribute on a model (except for a few of the ActiveRecord core attributes) is updatable via update_attributes
.
If you want to protect attributes from being updated you either need to single them out using attr_protected
or you can trigger whitelisting on the model by declaring at least one attribute attr_accessible
.
The initializer switches this round and makes whitelisting the default setting. With the intializer switched on, update_attributes
will only update attributes on your models which are declared attr_accessible
.
NB this is also true for .new
and .create
This strength and vulnerability in Rails is referred to as Mass Assignment and can happen when you pass a hash of arguments into either .new
or .create
as well as .update_attributes
. The latter tends to be the most vulnerable but all are susceptible.
There is good advice including an even cleaner solution than the one below on the official Rails documentation for protecting against mass assignment attacks.
Why this is needed
Take a simple User model:
create_users.rb migration:
class CreateUsers < ActiveRecord::Migration
def change
create_table :users do |t|
t.string :role
t.string :name
t.timestamps
end
end
end
and a very simple User class:
class User < ActiveRecord::Base
end
Why the User class is vulnerable
> u = User.create name: ‘Peter Nixey’, role: :subscriber;
=> #<User id: 1, role: :subscriber, name: "Peter Nixey",
created_at: "2012-03-05 09:39:31", updated_at: "2012-03-05 09:39:31">
By default, update_attributes
(which is what you’ll probably use in your update method) updates any attributes that are passed into it - usually via params[:model_name]
. It’s wonderfully quick and simple but open to abuse:
update_params
will for instance happily update not only your name but also your role:
> u.update_attributes name: ‘Jenson Button’, role: :superadmin;
=> #<User id: 1, role: "superadmin", name: "Jenson Button",
created_at: "2012-03-05 09:39:31", updated_at: "2012-03-05 09:40:53">
By fiddling with the user update form we just updated our role from subscriber to superadmin.
This is not good.
Since, by default, update_attributes
will update any parameter that’s passed into it, Homakov realised be could could use it to switch an SSH key for his own account to being one of the list of keys associated with one of the Rails GitHub account members.
Homakov assumed (correctly) that GitHub had a table containing users’ public keys. Each key has a value
and a user_id
. Homakov also correctly postulated that he might be able to update his own public key to have the user_id
of one of the Rails GitHub account members.
schematic of what the GitHub PublicKey#update method might look like:
class PublicKeyController < ApplicationController
before_filter :authorize_user
...
def update
@current_key = PublicKey.find_by_id params[:key]['id']
@current_key.update_attributes(params[:key])
end
end
Homakov PUT an update to his own existing public key which included a new user_id
. The user_id he used was that of a member of the Rails repository members.
The controller then simply updated all of the parameters Homakov passed it, including the new user_id
. With an SSH key on his machine registered to the repository of a Rails member all he then needed to do was push. This was the same hack he used for posting from the future.
Why don't I just avoid update_attributes
?
Why all the fuss about update_attributes
? If it’s so insecure why use it, why not manually update stuff using code such as
user.name = params[:user][‘name’]
This way everything would only be updated if we specifically updated it.
We could do but it would take five lines where update_attributes only takes one line. update_attributes
is also for better or worse, the Rails Way and so it’s a good idea to understand why it’s vulnerable and how to secure it.
How to protect update_attributes: attr_protected (not recommended)
Everything that happened happened because the user_id
attribute should not have been updatable via update_attributes. Rails has a method to prevent exactly this and it’s called attr_protected
.
class User < ActiveRecord::Base
attr_protected :role
end
With that line added it doesn’t matter whether we pass the role in via a PUT, it still won’t update:
u = User.create name: "Peter Nixey", role: :subscriber;
u.update_attributes role: :superadmin
WARNING: Can't mass-assign protected attributes: role
=> true
The problem is that attr_protected
only protects us on attributes we actually declare to be attr_protected
. It only works when we remember to add it. If we don’t put it in we don’t get protection.
I prefer to know I’m protected and safe until I chose to be unsafe and that (in theory) is what attr_accessible
gives us.
A bit safer protection: attr_accessible
attr_accessible
is the recommended method of tackling this problem. It’s actually a little bit of a misnomer since it's less about making an attribute accessible (it already was) and more about making it inaccessible.
Delcaring any attribute as attr_accessible
implies that all the other attributes are not accessible. Think of its real value less as being attr_accessible
and more as being attr_whitelist
class User < ActiveRecord::Base
attr_accessible :name
end
:role
is now protected since we haven't declared it attr_accessible
:
u = User.create name: "Peter Nixey", role: :subscriber;
u.update_attributes role: :superadmin
WARNING: Can't mass-assign protected attributes: role
=> true
The nice thing about attr_accessible
is that all new attributes are protected by default. If we add an account type to our database
class AddAccountType < ActiveRecord::Migration
def change
add_column :users, :account_type, :string
end
end
and leave our model unchanged:
class User < ActiveRecord::Base
attr_accessible :name
end
then the account_type
field is automatically protected:
u = User.create name: “Peter Nixey”, account_type: ‘free’;
u.update_attributes account_type: ‘paid’
WARNING: Can't mass-assign protected attributes: account_type
=> true
Either way you can still manually update attributes
We've not locked ourselves out of our own model. We can still update role directly it’s simply that it’s not vulnerable to being injected during update_attributes
u.role = :superadmin
u.save
(0.2ms) UPDATE "users" SET "role" = 'superadmin',
"updated_at" = '2012-03-05 10:11:05.042023' WHERE "users"."id" = 1
=> true
However, even attr_accessible only protects us when we remember it
The problem with attr_protected
was that it only protected us when we remembered to add it to the attribute.
The problem with attr_accessible
is that it only protects us when we remember to add it to the model. Sometimes (as GitHub showed us) it’s easy to forget to do that.
###The disable_mass_assignment initializer protects us by default
Create a new file:
config/initializers/disable_mass_assignment.rb
ActiveRecord::Base.send(:attr_accessible, nil)
The beauty of this is that it effectively adds attr_accessible
to every model we create (actually what it does it take it away by default but it comes to the same thing). No attribute can be updated unless we declare it attr_accessible
. We’re secure until we decide otherwise.
NB everything described here also applies to .create
and .new
Althugh the GitHub hack was done using update_attributes
, almost everything that's described here is also true for .create
and .new
too.
If you are not careful with the inputs that you feed into your object creation methods they too will initialize attributes on the object which you might not want to be initalized. For instance User.create( params[:user] ) woud also be vulnerable to role: :superadmin
being passed as a parameter.
The ability of Rails objects to assign multiple variables simultaneously is referred to as mass assignment (docs).
Possible issues you might have with the initializer
Once you setup the initializer, the first thing you’re going to need to do is declare all relevant attributes as attr_accessible
.
A good test suite will help a lot here but either way you need to go through each model adding each parameter that you want to be accessible to your attr_accessible
arguments:
class User
attr_accessible :email, :first_name, :last_name, :full_name
end
You’re going to have some frustrations. There are going to be things that you don’t see coming which will fail silently. Problems I’ve had are:
- Authlogic: you need to remember to make attributes like password, email etc accessible
- Paperclip: remember to make paperclip attributes accessible
- Nested attributes: Instructions here
- Delayed Job (fix from @borski): add the following line to make the appropriate delayed_Job attributes work:
Delayed::Job.attr_accessible :priority, :payload_object, :run_at, :locked_at, :failed_at, :locked_by
I’m sure you’ll hit other issues too but you can generally knock them off by adding attributes one by one to the list of accessible ones.
###How Rails could address this
I wouldn’t pretend to have anything like the oversight of the Rails landscape that the Rails core team do. I’ve only built a very few apps and I’m no guru. However...
The argument has been made several times that it is up to the app builder to secure their own app. I don't agree with this though. Rails’ mantra is convention over configuration.
If the Rails team are going to stand by the mantra then they also need to accept that the Rails Way to handle updates is conventionally insecure until configured otherwise.
Enforcing that attributes have to be declared attr_accessible
by default would immediately make things better.
What should the default authorization setting to be, on or off?
Yehuda Katz makes the point that this is an authorization issue which is not a framework issue. I think the question here though is not “where should authorization be handled” but “what should the default setting for authorization be”.
In most other public-facing interfaces in Rails the default setting for authorization is unauthorized. You can’t even reach a controller method unless you explicitly create a route for it. update_attributes
however defaults to authorized.
There are a lot of sites vulnerable and more being built every day
Rails is a brilliant framework designed by brilliant contributors. One of the reasons I like coding in it is that I feel that I always learn from the code I find (in PHP I generally wanted to rewrite it). This one weakness has always bugged me though and I feel it doesn't do Rails justice.
If GitHub, one of the best Rails teams on the planet can be taken out so easily, in so many places by such a simple hack then there is a real and present issue. As Yehuda says, not all security vulnerabilities can be fixed by the framework however this one can and it would make sense to do so.
Author: Peter Nixey
Twitter: http://twitter.com/peternixey
Blog: http://peternixey.com
###Update - an update has been added to Rails to whitelist by default
6th March 2012
Part a result of everything that happened in the past few days, new projects in Rails will default to whitelisting all attribute arguments by default:
https://github.com/rails/rails/commit/06a3a8a458e70c1b6531ac53c57a302b162fd736
This is a slightly different fix than the one described here and almost certainly better - I would advise using the official Rails solution which simply involves a change in your config file:
config/application.rb
...
config.active_record.whitelist_attributes = true
It will not affect existing projects which will still need to address their issues independently however it will mean that new Rails projects will be required to whitelist attributes by default. The commit is planned to be included in Rails 3-2-stable
Further reading:
@petenixey: Thanks Pete for the reply. Just to make sure I understand things completely, the section of code:
has as an intended use the ability for a signed-in user to update his own public key. The filter :authorize_user is only verifying that the user is signed in. However, a malicious, signed in user can replace his own user id in the params hash with an id of whomever he would like and then the line:
will return the key of the id that the malicious user specified. He can then update this key with his own public key, giving him push ability to a repo he would normally not have access to.
Please correct me if I'm wrong in this reasoning. Also, sorry to be reiterating what you stated in the gist above, a lot of this is new to me.
Finally, I've been following the rather excellent ruby on rails tutorial by Michael Hartl and in chapter 9 it deals with user authentication and authorization (obviously I have a lot to learn on this subject). Would an alternative solution to the GitHub vulnerability be something along the line of section 9.15 in the tutorial (http://ruby.railstutorial.org/chapters/updating-showing-and-deleting-users?version=3.2#code:correct_user_before_filter) where the :authorize_user filter in your example is replaced by:
thanks,
nate