The Universe of Discourse
Sat, 12 Dec 2009

On failing open
An axiom of security analysis is that nearly all security mechanisms must fail closed. What this means is that if there is an uncertainty about whether to grant or to deny access, the right choice is nearly always to deny access.

For example, consider a login component that accepts a username and a password and then queries a remote authentication server to find out if the password is correct. If the connection to the authentication server fails, or if the authentication server is down, the login component must decide whether to grant or deny access, in the absence of correct information from the server. The correct design is almost certainly to "fail closed", and to deny access.

I used to teach security classes, and I would point out that programs sometimes have bugs, and do the wrong thing. If your program has failed closed, and if this is a bug, then you have an irate user. The user might call you up and chew you out, or might chew you out to your boss, and they might even miss a crucial deadline because your software denied them access when it should have granted access. But these are relatively small problems. If your program has failed open, and if this is a bug, then the user might abscond with the entire payroll and flee to Brazil.

(I was once teaching one of these classes in Lisbon, and I reached the "flee to Brazil" example without having realized ahead of time that this had greater potential to offend the Portuguese than many other people. So I apologized. But my hosts very kindly told me that they would have put it the same way, and that in fact the Mayor of Lisbon had done precisely what I described a few years before. The moral of the story is to read over the slides ahead of time before giving the talk.)

But I digress. One can find many examples in the history of security that failed the wrong way.

However, the issue is on my mind because I was at a job interview a few weeks ago with giant media corporation XYZ. At the interview, we spent about an hour talking about an architectural problem they were trying to solve. XYZ operates a web site where people can watch movies and TV programs online. Thy would like to extend the service so that people who subscribe to premium cable services, such as HBO, can authenticate themselves to the web site and watch HBO programs there; HBO non-subscribers should get only free TV content. The problem in this case was that the authentication data was held on an underpowered legacy system that could serve only a small fraction of the requests that came in.

The solution was to cache the authentication data on a better system, and gather and merge change information from the slow legacy system as possible.

I observed during the discussion that this was a striking example of the rare situation in which one wants the authentication system to fail open instead of closed. For suppose one grants access that should not be granted. Then someone on the Internet gets to watch a movie or an episode of The Sopranos for free, which is not worth getting excited about and which happens a zillion times a day anyhow.

But suppose the software denies access that should have been granted. Then there is a legitimate paying customer who has paid to watch The Sopranos, and we told them no. Now they are a legitimately irate customer, which is bad, and they may call the support desk, costing XYZ Corp a significant amount of money, which is also bad. So all other things being equal, one should err on the side of lenity: when in doubt, grant access.

I would like to thank Andrew Lenards for his gift.

[Other articles in category /tech] permanent link